Enhance the verification of the logic that involves the use of think_tool

.env.example

@@ -10,4 +10,7 @@ LANGSMITH_TRACING=
 SUPABASE_KEY=
 SUPABASE_URL=
 # Should be set to true for a production deployment on Open Agent Platform. Should be set to false otherwise, such as for local development.
-GET_API_KEYS_FROM_CONFIG=false
+GET_API_KEYS_FROM_CONFIG=false
+
+# Setting it to true ensures that the think_tool tool can execute correctly. Otherwise, there is a very small probability that it will not be executed.
+FORCE_THINK_TOOL=False

src/open_deep_research/configuration.py

@@ -39,6 +39,16 @@ class Configuration(BaseModel):
     """Main configuration class for the Deep Research agent."""
 
     # General Configuration
+    force_think_tool: bool = Field(
+        default=False,
+        metadata={
+            "x_oap_ui_config": {
+                "type": "boolean",
+                "default": False,
+                "description": "Require agents to call think_tool before executing other tools"
+            }
+        }
+    )
     max_structured_output_retries: int = Field(
         default=3,
         metadata={

src/open_deep_research/deep_researcher.py

@@ -27,6 +27,10 @@
     lead_researcher_prompt,
     research_system_prompt,
     transform_messages_into_research_topic_prompt,
+    force_think_tool_supervisor_before_conduct_research_reminder,
+    force_think_tool_supervisor_after_conduct_research_reminder,
+    force_think_tool_researcher_after_call_search_tool_reminder,
+    has_mixed_think_tool_calls_reminder,
 )
 from open_deep_research.state import (
     AgentInputState,
@@ -50,6 +54,8 @@
     openai_websearch_called,
     remove_up_to_last_ai_message,
     think_tool,
+    has_mixed_think_tool_calls,
+    get_previous_tool_name,
 )
 
 # Initialize a configurable model that we will use throughout the agent
@@ -260,6 +266,17 @@ async def supervisor_tools(state: SupervisorState, config: RunnableConfig) -> Co
                 "research_brief": state.get("research_brief", "")
             }
         )
+
+    # Check if think_tool is mixed with other tools
+    if has_mixed_think_tool_calls(most_recent_message.tool_calls):
+        return Command(
+            goto="supervisor",
+            update={
+                "supervisor_messages": [
+                    HumanMessage(content=has_mixed_think_tool_calls_reminder)
+                ]
+            }
+        )
 
     # Step 2: Process all tool calls together (both think_tool and ConductResearch)
     all_tool_messages = []
@@ -285,6 +302,33 @@ async def supervisor_tools(state: SupervisorState, config: RunnableConfig) -> Co
         if tool_call["name"] == "ConductResearch"
     ]
 
+    # Check if think_tool is called correctly before calling ConductResearch
+    if configurable.force_think_tool:
+        previous_tool_name = get_previous_tool_name(supervisor_messages)
+
+        # Ensure that think_tool is called correctly before calling ConductResearch
+        if conduct_research_calls:
+            if previous_tool_name != "think_tool":
+                return Command(
+                    goto="supervisor",
+                    update={
+                        "supervisor_messages": [
+                            HumanMessage(content=force_think_tool_supervisor_before_conduct_research_reminder)
+                        ]
+                    }
+                )
+
+        # Ensure that think_tool is called correctly after calling ConductResearch
+        if previous_tool_name == "ConductResearch" and most_recent_message.tool_calls[0]["name"] != "think_tool":
+            return Command(
+                goto="supervisor",
+                update={
+                    "supervisor_messages": [
+                        HumanMessage(content=force_think_tool_supervisor_after_conduct_research_reminder)
+                    ]
+                }
+            )
+
     if conduct_research_calls:
         try:
             # Limit concurrent research units to prevent resource exhaustion
@@ -463,6 +507,17 @@ async def researcher_tools(state: ResearcherState, config: RunnableConfig) -> Co
     if not has_tool_calls and not has_native_search:
         return Command(goto="compress_research")
 
+    # Check if think_tool is mixed with other tools
+    if has_mixed_think_tool_calls(most_recent_message.tool_calls):
+        return Command(
+            goto="researcher",
+            update={
+                "researcher_messages": [
+                    HumanMessage(content=has_mixed_think_tool_calls_reminder)
+                ],
+            }
+        )
+
     # Step 2: Handle other tool calls (search, MCP tools, etc.)
     tools = await get_all_tools(config)
     tools_by_name = {
@@ -472,6 +527,22 @@ async def researcher_tools(state: ResearcherState, config: RunnableConfig) -> Co
 
     # Execute all tool calls in parallel
     tool_calls = most_recent_message.tool_calls
+
+    # Check if think_tool is called correctly
+    if configurable.force_think_tool:
+
+        previous_tool_name = get_previous_tool_name(researcher_messages)
+
+        # Ensure that think_tool is called correctly after calling other tools
+        if previous_tool_name and previous_tool_name != "think_tool" and tool_calls[0]["name"] != "think_tool":
+            return Command(
+                goto="researcher",
+                update={
+                    "researcher_messages": [
+                        HumanMessage(content=force_think_tool_researcher_after_call_search_tool_reminder)
+                    ]
+                }
+            )
     tool_execution_tasks = [
         execute_tool_safely(tools_by_name[tool_call["name"]], tool_call["args"], config) 
         for tool_call in tool_calls

src/open_deep_research/prompts.py

@@ -365,4 +365,25 @@
 Remember, your goal is to create a summary that can be easily understood and utilized by a downstream research agent while preserving the most critical information from the original webpage.
 
 Today's date is {date}.
-"""
+"""
+
+
+
+force_think_tool_supervisor_before_conduct_research_reminder = (
+    "Before delegating any research with ConductResearch you must first call think_tool to "
+    "plan your approach. Reflect with think_tool, then decide whether to delegate."
+)
+
+force_think_tool_supervisor_after_conduct_research_reminder = (
+    "After calling ConductResearch, you must call think_tool to reflect on the results. "
+    "Use think_tool now to analyze progress and plan the next action."
+)
+
+force_think_tool_researcher_after_call_search_tool_reminder = (
+    "You must call think_tool to reflect on the latest findings before using other tools. "
+    "Use think_tool now to analyze progress and plan the next action."
+)
+
+has_mixed_think_tool_calls_reminder = (
+    "You cannot call think_tool in parallel with other tools. Please call think_tool first, then call the other tools."
+)

src/open_deep_research/utils.py

@@ -14,6 +14,7 @@
     AIMessage,
     HumanMessage,
     MessageLikeRepresentation,
+    ToolMessage,
     filter_messages,
 )
 from langchain_core.runnables import RunnableConfig
@@ -923,3 +924,20 @@ def get_tavily_api_key(config: RunnableConfig):
         return api_keys.get("TAVILY_API_KEY")
     else:
         return os.getenv("TAVILY_API_KEY")
+
+
+def has_mixed_think_tool_calls(tool_calls: list[dict]) -> bool:
+    """Return True if think_tool is mixed with other tools in the same turn."""
+    has_think = any(call["name"] == "think_tool" for call in tool_calls)
+    has_other = any(call["name"] != "think_tool" for call in tool_calls)
+    return has_think and has_other
+
+def get_previous_tool_name(messages: list[object]) -> str | None:
+    """Return the name of the most recent tool message before the latest message. If there is no previous tool message, return None."""
+    # Skip the most recent message (typically the current AI reply with tool calls)
+    for i in range(len(messages) - 1, -1, -1):
+        if isinstance(messages[i], ToolMessage):
+            return messages[i].name
+        if isinstance(messages[i], dict) and messages[i].get("type") == "tool":
+            return messages[i].get("name")
+    return None