Merge branch 'main' into dependabot/pip/models/gemini/protobuf-6.33.5

Author: crazywoola

.gitignore

@@ -188,3 +188,8 @@ pyrightconfig.json
 mypy.ini
 tox.ini
 *.difypkg
+models/openai_api_compatible/PR_DESCRIPTION.md
+
+# local venv
+models/volcengine/.venv/
+models/volcengine/uv.lock

agent-strategies/cot_agent/manifest.yaml

@@ -1,4 +1,4 @@
-version: 0.0.31
+version: 0.0.32
 type: plugin
 author: "langgenius"
 name: "agent"

agent-strategies/cot_agent/output_parser/cot_output_parser.py

@@ -7,6 +7,9 @@
 from dify_plugin.interfaces.agent import AgentScratchpadUnit
 
 PREFIX_DELIMITERS = frozenset({"\n", " ", ""})
+# Tags injected by Gemini when include_thoughts=True; stripped so ReAct sees only Thought:/Action:/FinalAnswer:
+THINK_START = "<think>"
+THINK_END = "</think>"
 
 
 class ReactState(Enum):
@@ -115,11 +118,63 @@ def step(self, delta: str) -> tuple[bool, ReactChunk | None, bool, bool]:
         answer_matcher = PrefixMatcher(ReactState.ANSWER)
         thought_matcher = PrefixMatcher(ReactState.THINKING)
 
+        _in_think = False
+        _think_buf = ""
+        _think_depth = 0
         for response in llm_response:
             if response.delta.usage:
                 usage_dict["usage"] = response.delta.usage
-            response_content = response.delta.message.content
-            if not isinstance(response_content, str):
+            raw = response.delta.message.content
+            if isinstance(raw, str):
+                response_content = raw
+            elif isinstance(raw, list):
+                # Plugins (e.g. Gemini) send content as list; some items may be non-text (e.g. image)
+                parts = [
+                    s
+                    for c in raw
+                    if isinstance(s := (getattr(c, "data", None) or getattr(c, "text", None)), str)
+                ]
+                response_content = "".join(parts)
+            else:
+                continue
+            if not response_content:
+                continue
+            # When include_thoughts=True, Gemini injects <think>...</think>; strip across chunks so
+            # ReAct parser only sees Thought:/Action:/FinalAnswer: from the model reply.
+            # Nested <think> tags are supported via a depth counter.
+            if THINK_START in response_content or THINK_END in response_content or _in_think:
+                buf = _think_buf + response_content
+                _think_buf = ""
+                out = []
+                i = 0
+                while i < len(buf):
+                    if _in_think:
+                        end_j = buf.find(THINK_END, i)
+                        start_j = buf.find(THINK_START, i)
+                        if end_j == -1 and start_j == -1:
+                            _think_buf = buf[i:]
+                            break
+                        if start_j != -1 and (end_j == -1 or start_j < end_j):
+                            _think_depth += 1
+                            i = start_j + len(THINK_START)
+                        else:
+                            j = end_j
+                            _think_depth -= 1
+                            if _think_depth <= 0:
+                                _in_think = False
+                                _think_depth = 0
+                            i = j + len(THINK_END)
+                    else:
+                        j = buf.find(THINK_START, i)
+                        if j == -1:
+                            out.append(buf[i:])
+                            break
+                        out.append(buf[i:j])
+                        _in_think = True
+                        _think_depth = 1
+                        i = j + len(THINK_START)
+                response_content = "".join(out)
+            if not response_content:
                 continue
 
             # stream

agent-strategies/cot_agent/strategies/ReAct.py

@@ -121,6 +121,7 @@ def _invoke(self, parameters: dict[str, Any]) -> Generator[AgentInvokeMessage]:
         run_agent_state = True
         llm_usage: dict[str, Optional[LLMUsage]] = {"usage": None}
         final_answer = ""
+        final_answer_already_streamed = False  # True when final_answer = thought (no action), already sent as THINKING
         prompt_messages: list[PromptMessage] = []
 
         # Init model
@@ -217,7 +218,9 @@ def _invoke(self, parameters: dict[str, Any]) -> Generator[AgentInvokeMessage]:
                     assert isinstance(react_chunk, ReactChunk)
                     chunk_state = react_chunk.state
                     chunk = react_chunk.content
-                    yield self.create_text_message(chunk)
+                    # Stream TEXT only for THINKING; ANSWER is sent once at end via final_answer
+                    if chunk_state != ReactState.ANSWER:
+                        yield self.create_text_message(chunk)
                     if chunk_state == ReactState.ANSWER:
                         final_answer += chunk
                     elif chunk_state == ReactState.THINKING:
@@ -266,6 +269,7 @@ def _invoke(self, parameters: dict[str, Any]) -> Generator[AgentInvokeMessage]:
             )
             if not scratchpad.action:
                 final_answer = scratchpad.thought
+                final_answer_already_streamed = True  # thought was already streamed as THINKING chunks
             else:
                 if scratchpad.action.action_name.lower() == "final answer":
                     try:
@@ -419,7 +423,8 @@ def _invoke(self, parameters: dict[str, Any]) -> Generator[AgentInvokeMessage]:
             )
             iteration_step += 1
 
-        # yield self.create_text_message(final_answer)
+        if not final_answer_already_streamed and final_answer:
+            yield self.create_text_message(final_answer)
 
         # If context is a list of dict, create retriever resource message
         if isinstance(react_params.context, list):

datasources/github/datasources/github.py

@@ -80,8 +80,8 @@ def _get_pages(self, datasource_parameters: dict[str, Any]) -> DatasourceGetPage
                 raise ValueError(
                     f"Invalid 'affiliation' parts: {', '.join(parts - allowed_affiliations)}. Allowed: {', '.join(allowed_affiliations)}.")
 
-        _type = datasource_parameters.get("type", "all")
-        if _type not in {"all", "owner", "public", "private", "member"}:
+        _type = datasource_parameters.get("type")
+        if _type and _type not in {"all", "owner", "public", "private", "member"}:
             raise ValueError(
                 f"Invalid 'type' parameter: {_type}. Allowed values are: all, owner, public, private, member.")
 

datasources/github/manifest.yaml

@@ -1,4 +1,4 @@
-version: 0.4.0
+version: 0.4.1
 type: plugin
 author: langgenius
 name: github_datasource

models/aihubmix/manifest.yaml

@@ -1,4 +1,4 @@
-version: 0.0.16
+version: 0.0.17
 type: plugin
 author: langgenius
 name: aihubmix

models/aihubmix/pyproject.toml

@@ -12,7 +12,7 @@ dependencies = [
     "openai~=2.3.0",
     "google-genai>=1.56.0",
     "anthropic~=0.57.1",
-    "pillow~=11.3.0",
+    "pillow~=12.1.1",
     "urllib3>=2.6.3",
 ]
 

models/aihubmix/requirements.txt

@@ -3,5 +3,5 @@ httpx>=0.28.1,<0.29.dev0
 openai~=2.3.0
 google-genai>=1.56.0
 anthropic~=0.57.1
-pillow~=11.3.0
+pillow~=12.1.1
 urllib3>=2.6.3