agenthub-python 0.4.8__tar.gz → 0.4.10__tar.gz

{agenthub_python-0.4.8 → agenthub_python-0.4.10}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: agenthub-python
-Version: 0.4.8
+Version: 0.4.10
 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-0.4.8 → agenthub_python-0.4.10}/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-0.4.8 → agenthub_python-0.4.10}/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):
@@ -65,13 +71,13 @@ 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
@@ -99,6 +105,12 @@ 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 +143,10 @@ 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, "
+                "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-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-0.4.8 → agenthub_python-0.4.10}/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-0.4.8 → agenthub_python-0.4.10}/agenthub/deepseek_v4/client.py

@@ -103,10 +103,17 @@ 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"]]
 

{agenthub_python-0.4.8/agenthub/gemini3_7 → agenthub_python-0.4.10/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-0.4.8/agenthub/gemini3_7 → agenthub_python-0.4.10/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-0.4.8 → agenthub_python-0.4.10}/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-0.4.8 → agenthub_python-0.4.10}/agenthub/gpt5_6/client.py

@@ -33,7 +33,7 @@ from ..types import (
     UniMessage,
     UsageMetadata,
 )
-from ..utils import is_debug_enabled
+from ..utils import is_debug_enabled, openai_image_detail
 
 
 class GPT5_6Client(LLMClient):
@@ -71,6 +71,14 @@ class GPT5_6Client(LLMClient):
             return {"mode": "required", "tools": [{"type": "function", "name": name} for name in tool_choice]}
         return tool_choice
 
+    def _convert_image_url(self, image_url: str) -> dict[str, str]:
+        """Convert an image URL to an input_image item, at the detail the API needs to read it."""
+        item = {"type": "input_image", "image_url": image_url}
+        if detail := openai_image_detail(self._model, image_url):
+            item["detail"] = detail
+
+        return item
+
     def transform_uni_config_to_model_config(self, config: UniConfig) -> dict[str, Any]:
         """
         Transform universal configuration to OpenAI Responses API configuration.
@@ -98,8 +106,12 @@ class GPT5_6Client(LLMClient):
 
         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"]]
@@ -158,7 +170,7 @@ class GPT5_6Client(LLMClient):
                     else:
                         content_items.append({"type": "output_text", "text": item["text"]})
                 elif item["type"] == "image_url":
-                    content_items.append({"type": "input_image", "image_url": item["image_url"]})
+                    content_items.append(self._convert_image_url(item["image_url"]))
                 elif item["type"] == "thinking":
                     # rebuild the reasoning item from the recorded wire fields: the thinking
                     # text goes back through the channel that carried it (histories recorded
@@ -196,7 +208,7 @@ class GPT5_6Client(LLMClient):
                     tool_result = [{"type": "input_text", "text": item["text"]}]
                     if "images" in item:
                         for image_url in item["images"]:
-                            tool_result.append({"type": "input_image", "image_url": image_url})
+                            tool_result.append(self._convert_image_url(image_url))
 
                     input_list.append(
                         {"type": "function_call_output", "call_id": item["tool_call_id"], "output": tool_result}

{agenthub_python-0.4.8 → agenthub_python-0.4.10}/agenthub/openai_chat/client.py

@@ -35,7 +35,7 @@ from ..types import (
     UniMessage,
     UsageMetadata,
 )
-from ..utils import fix_openrouter_usage_metadata
+from ..utils import fix_openrouter_usage_metadata, openai_image_detail
 
 
 class OpenaiChatClient(LLMClient):
@@ -88,6 +88,14 @@ class OpenaiChatClient(LLMClient):
 
         return tool_choice
 
+    def _convert_image_url(self, data_url: str) -> dict[str, Any]:
+        """Convert a fetched image to an image_url part, at the detail the API needs to read it."""
+        image_url = {"url": data_url}
+        if detail := openai_image_detail(self._model, data_url):
+            image_url["detail"] = detail
+
+        return {"type": "image_url", "image_url": image_url}
+
     def transform_uni_config_to_model_config(self, config: UniConfig) -> dict[str, Any]:
         """
         Transform universal configuration to OpenAI Chat Completions configuration.
@@ -146,7 +154,7 @@ class OpenaiChatClient(LLMClient):
                     content_parts.append({"type": "text", "text": item["text"]})
                 elif item["type"] == "image_url":
                     base64_image = await self._convert_image_url_to_base64(item["image_url"])
-                    content_parts.append({"type": "image_url", "image_url": {"url": base64_image}})
+                    content_parts.append(self._convert_image_url(base64_image))
                 elif item["type"] == "thinking":
                     thinking += item["thinking"]
                     thinking_fields.add((item.get("fidelity") or {}).get("reasoning_field"))
@@ -169,12 +177,12 @@ class OpenaiChatClient(LLMClient):
 
                     if "images" in item and item["images"]:
                         for image_url in item["images"]:
-                            base64_image = await self._convert_image_url_to_base64(image_url)
+                            part = self._convert_image_url(await self._convert_image_url_to_base64(image_url))
                             if "siliconflow.cn" in str(self._client.base_url):
                                 # siliconflow does not support image_url in tool result
-                                content_parts.append({"type": "image_url", "image_url": {"url": base64_image}})
+                                content_parts.append(part)
                             else:
-                                content.append({"type": "image_url", "image_url": {"url": base64_image}})
+                                content.append(part)
 
                     # Tool results are sent as separate messages
                     openai_messages.append(

agenthub_python-0.4.10/agenthub/openai_chat_vllm_adapter/__init__.py

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

agenthub_python-0.4.10/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-0.4.8 → agenthub_python-0.4.10}/agenthub/openai_responses/client.py

@@ -33,7 +33,7 @@ from ..types import (
     UniMessage,
     UsageMetadata,
 )
-from ..utils import is_debug_enabled
+from ..utils import is_debug_enabled, openai_image_detail
 
 
 class OpenaiResponsesClient(LLMClient):
@@ -71,6 +71,14 @@ class OpenaiResponsesClient(LLMClient):
             return {"mode": "required", "tools": [{"type": "function", "name": name} for name in tool_choice]}
         return tool_choice
 
+    def _convert_image_url(self, image_url: str) -> dict[str, str]:
+        """Convert an image URL to an input_image item, at the detail the API needs to read it."""
+        item = {"type": "input_image", "image_url": image_url}
+        if detail := openai_image_detail(self._model, image_url):
+            item["detail"] = detail
+
+        return item
+
     def transform_uni_config_to_model_config(self, config: UniConfig) -> dict[str, Any]:
         """
         Transform universal configuration to OpenAI Responses-compatible configuration.
@@ -92,6 +100,11 @@ 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"):
@@ -156,7 +169,7 @@ class OpenaiResponsesClient(LLMClient):
                     else:
                         content_items.append({"type": "output_text", "text": item["text"]})
                 elif item["type"] == "image_url":
-                    content_items.append({"type": "input_image", "image_url": item["image_url"]})
+                    content_items.append(self._convert_image_url(item["image_url"]))
                 elif item["type"] == "thinking":
                     # the wire shape differs by server: OpenAI-style servers stream summaries and
                     # demand the summary key back (with encrypted_content preserved), while
@@ -192,7 +205,7 @@ class OpenaiResponsesClient(LLMClient):
                     tool_result = [{"type": "input_text", "text": item["text"]}]
                     if "images" in item:
                         for image_url in item["images"]:
-                            tool_result.append({"type": "input_image", "image_url": image_url})
+                            tool_result.append(self._convert_image_url(image_url))
 
                     input_list.append(
                         {"type": "function_call_output", "call_id": item["tool_call_id"], "output": tool_result}

{agenthub_python-0.4.8 → agenthub_python-0.4.10}/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"],
     },

agenthub_python-0.4.10/agenthub/utils.py

@@ -0,0 +1,243 @@
+# 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.
+
+import base64
+import math
+import os
+import re
+from typing import Literal
+
+from .types import UsageMetadata
+
+
+def fix_openrouter_usage_metadata(usage_metadata: UsageMetadata, base_url: str) -> UsageMetadata:
+    """
+    Fix the usage metadata for OpenRouter.
+
+    OpenRouter occasionally does not include the reasoning tokens to the completion tokens.
+
+    Args:
+        usage_metadata (UsageMetadata): The usage metadata.
+        base_url (str): The API URL.
+
+    Returns:
+        UsageMetadata: The fixed usage metadata.
+    """
+    fixed_usage_metadata = usage_metadata.copy()
+    if "openrouter.ai" in base_url and fixed_usage_metadata["response_tokens"] < 0:
+        fixed_usage_metadata["response_tokens"] += fixed_usage_metadata["thoughts_tokens"] or 0
+
+    return fixed_usage_metadata
+
+
+def is_debug_enabled() -> bool:
+    """
+    Whether AGENTHUB_DEBUG asks the clients to fail loudly on output they do not recognize.
+
+    Streaming clients skip an unrecognized event so that a gateway's own frames cannot kill a
+    long generation. The same silence hides a genuinely new provider event, so the guards stay
+    one environment variable away.
+
+    Returns:
+        bool: Whether debug mode is on.
+    """
+    return os.getenv("AGENTHUB_DEBUG", "").strip().lower() not in ("", "0", "false", "no", "off")
+
+
+def image_dimensions(data: bytes) -> tuple[int, int] | None:
+    """
+    Read the pixel dimensions from the header of a PNG, JPEG, GIF or WebP image.
+
+    Only the header is inspected, so the bytes may be a prefix of the file; a prefix that ends
+    before the dimensions are reached reads as unrecognized.
+
+    Args:
+        data (bytes): The image bytes, or a prefix of them.
+
+    Returns:
+        tuple[int, int] | None: The width and height, or None when the bytes are not a recognized image.
+    """
+    # PNG: an 8-byte signature, then the IHDR chunk with width and height
+    if len(data) >= 24 and data[:8] == b"\x89PNG\r\n\x1a\n" and data[12:16] == b"IHDR":
+        return int.from_bytes(data[16:20], "big"), int.from_bytes(data[20:24], "big")
+
+    # GIF: the logical screen size follows the 6-byte signature
+    if len(data) >= 10 and data[:6] in (b"GIF87a", b"GIF89a"):
+        return int.from_bytes(data[6:8], "little"), int.from_bytes(data[8:10], "little")
+
+    # WebP: a RIFF container whose first chunk names the bitstream flavour
+    if len(data) >= 30 and data[:4] == b"RIFF" and data[8:12] == b"WEBP":
+        chunk = data[12:16]
+        if chunk == b"VP8 " and data[23:26] == b"\x9d\x01\x2a":
+            # lossy: a 3-byte frame tag and the key frame start code precede the size, whose top
+            # two bits are a scaling hint
+            width = int.from_bytes(data[26:28], "little") & 0x3FFF
+            height = int.from_bytes(data[28:30], "little") & 0x3FFF
+            return width, height
+        if chunk == b"VP8L" and data[20] == 0x2F:
+            # lossless: 14 bits of width minus one, then 14 bits of height minus one
+            bits = int.from_bytes(data[21:25], "little")
+            return (bits & 0x3FFF) + 1, ((bits >> 14) & 0x3FFF) + 1
+        if chunk == b"VP8X":
+            # extended: the canvas size minus one, 24 bits each, after the flags
+            return int.from_bytes(data[24:27], "little") + 1, int.from_bytes(data[27:30], "little") + 1
+        return None
+
+    # JPEG: walk the marker segments to the first frame header (SOFn)
+    if len(data) >= 4 and data[:2] == b"\xff\xd8":
+        offset = 2
+        while offset + 3 < len(data):
+            if data[offset] != 0xFF:
+                return None
+            marker = data[offset + 1]
+            if marker == 0xFF:  # fill byte ahead of a marker
+                offset += 1
+                continue
+            if marker in (0x01, 0xD8) or 0xD0 <= marker <= 0xD7:  # TEM, SOI and RSTn stand alone, without a length
+                offset += 2
+                continue
+            if marker in (0xD9, 0xDA):  # end of image, or scan data before any frame header
+                return None
+            if 0xC0 <= marker <= 0xCF and marker not in (0xC4, 0xC8, 0xCC):
+                if offset + 8 >= len(data):
+                    return None
+                # length, precision, then height before width
+                height = int.from_bytes(data[offset + 5 : offset + 7], "big")
+                width = int.from_bytes(data[offset + 7 : offset + 9], "big")
+                return width, height
+            offset += 2 + int.from_bytes(data[offset + 2 : offset + 4], "big")
+        return None
+
+    return None
+
+
+# The patch count above which the OpenAI vision API rejects an image instead of resizing it
+# (Images and vision guide, "Choose an image detail level":
+# https://developers.openai.com/api/docs/guides/images-vision).
+_OPENAI_PATCH_LIMIT = 30000
+
+# The longest side the OpenAI vision API keeps at `original` detail; a larger image is scaled
+# down to fit it before the patches are counted (the same guide, model sizing table).
+_OPENAI_ORIGINAL_MAX_SIDE = 65535
+
+# Base64 characters decoded first: 48 bytes, enough for any PNG, GIF or WebP header. A JPEG's
+# frame header may sit behind metadata segments, so its window grows by the factor below until
+# the header is found or the payload runs out.
+_HEADER_PROBE_CHARS = 64
+_HEADER_WINDOW_GROWTH = 4
+
+# Characters outside the base64 alphabet, URL-safe variant included, and the map from that
+# variant back to the standard alphabet.
+_NOT_BASE64 = re.compile(r"[^A-Za-z0-9+/_-]")
+_URL_SAFE_TO_STANDARD = str.maketrans("-_", "+/")
+
+
+def _base64_payload_start(data_url: str) -> int:
+    """
+    Locate the payload of a base64 data URL.
+
+    Args:
+        data_url (str): The URL.
+
+    Returns:
+        int: The index of the payload's first character, or -1 when the URL is not a base64 data URL.
+    """
+    if not data_url.startswith("data:"):
+        return -1
+    comma = data_url.find(",")
+    if comma < 0:
+        return -1
+    # the token is case-insensitive and may follow a space, as a browser reads it
+    params = [param.strip().lower() for param in data_url[5:comma].split(";")]
+    return comma + 1 if "base64" in params else -1
+
+
+def _decode_base64_prefix(text: str) -> bytes:
+    """
+    Decode a prefix of a base64 payload the way Node's Buffer does.
+
+    Whitespace and other stray characters are skipped, the URL-safe alphabet is accepted, and a
+    cut inside a 4-character group or missing padding yields the bytes that are complete, so the
+    Python and TypeScript clients measure the same image alike.
+
+    Args:
+        text (str): The leading characters of the payload.
+
+    Returns:
+        bytes: The decoded bytes.
+    """
+    chars = _NOT_BASE64.sub("", text).translate(_URL_SAFE_TO_STANDARD)
+    if len(chars) % 4 == 1:  # a lone trailing character carries no whole byte
+        chars = chars[:-1]
+    return base64.b64decode(chars + "=" * (-len(chars) % 4))
+
+
+def exceeds_openai_patch_limit(image_url: str) -> bool:
+    """
+    Whether the OpenAI vision API would reject an image at `original` detail.
+
+    The API covers an image with 32-pixel patches and rejects one that needs more than 30,000 of
+    them after its own resizing; at `original` detail the only resizing is the 65,535-pixel cap
+    on either side. Only a base64 data URL can be measured here: an HTTP(S) URL answers False.
+    The Responses clients pass such a URL through for the API to fetch; the Chat client fetches
+    it into a data URL first, so it measures the fetched bytes.
+
+    Args:
+        image_url (str): The image URL.
+
+    Returns:
+        bool: Whether the API would reject the image.
+    """
+    start = _base64_payload_start(image_url)
+    if start < 0:
+        return False
+    # decode a growing prefix rather than the whole payload
+    chars = _HEADER_PROBE_CHARS
+    while True:
+        data = _decode_base64_prefix(image_url[start : start + chars])
+        size = image_dimensions(data)
+        if size is not None or data[:2] != b"\xff\xd8" or start + chars >= len(image_url):
+            break
+        chars *= _HEADER_WINDOW_GROWTH
+    if size is None:
+        return False
+
+    width, height = size
+    longest = max(width, height)
+    if longest > _OPENAI_ORIGINAL_MAX_SIDE:
+        width = math.floor(width * _OPENAI_ORIGINAL_MAX_SIDE / longest + 0.5)
+        height = math.floor(height * _OPENAI_ORIGINAL_MAX_SIDE / longest + 0.5)
+    return math.ceil(width / 32) * math.ceil(height / 32) > _OPENAI_PATCH_LIMIT
+
+
+def openai_image_detail(model: str, image_url: str) -> Literal["high"] | None:
+    """
+    The `detail` an OpenAI image part needs so that the API reads the image.
+
+    GPT-5.6 reads the default `auto` detail as `original`, which keeps the image's own
+    dimensions and rejects one over 30,000 patches instead of resizing it; `high` has the API
+    fit it into 2,500 patches, so the image is read instead of refused. Every other model keeps
+    a patch budget at every detail level, so no other model gets the field.
+
+    Args:
+        model (str): The model id the request is sent with.
+        image_url (str): The image URL as it goes on the wire.
+
+    Returns:
+        Literal["high"] | None: "high" when the image needs it, otherwise None.
+    """
+    if "gpt-5.6" in model.lower() and exceeds_openai_patch_limit(image_url):
+        return "high"
+
+    return None

{agenthub_python-0.4.8 → agenthub_python-0.4.10}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "agenthub-python"
-version = "0.4.8"
+version = "0.4.10"
 description = "AgentHub is the LLM API Hub for the Agent era, built for high-precision autonomous agents."
 keywords = [
     "agent",

{agenthub_python-0.4.8 → agenthub_python-0.4.10}/pyproject.toml.orig

@@ -1,6 +1,6 @@
 [project]
 name = "agenthub-python"
-version = "0.4.8"
+version = "0.4.10"
 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"

agenthub_python-0.4.8/agenthub/utils.py

@@ -1,51 +0,0 @@
-# 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.
-
-import os
-
-from .types import UsageMetadata
-
-
-def fix_openrouter_usage_metadata(usage_metadata: UsageMetadata, base_url: str) -> UsageMetadata:
-    """
-    Fix the usage metadata for OpenRouter.
-
-    OpenRouter occasionally does not include the reasoning tokens to the completion tokens.
-
-    Args:
-        usage_metadata (UsageMetadata): The usage metadata.
-        base_url (str): The API URL.
-
-    Returns:
-        UsageMetadata: The fixed usage metadata.
-    """
-    fixed_usage_metadata = usage_metadata.copy()
-    if "openrouter.ai" in base_url and fixed_usage_metadata["response_tokens"] < 0:
-        fixed_usage_metadata["response_tokens"] += fixed_usage_metadata["thoughts_tokens"] or 0
-
-    return fixed_usage_metadata
-
-
-def is_debug_enabled() -> bool:
-    """
-    Whether AGENTHUB_DEBUG asks the clients to fail loudly on output they do not recognize.
-
-    Streaming clients skip an unrecognized event so that a gateway's own frames cannot kill a
-    long generation. The same silence hides a genuinely new provider event, so the guards stay
-    one environment variable away.
-
-    Returns:
-        bool: Whether debug mode is on.
-    """
-    return os.getenv("AGENTHUB_DEBUG", "").strip().lower() not in ("", "0", "false", "no", "off")