feat(tool_search): add deferred tool loading support for GPT-5.4

crates/forge_app/src/app.rs

@@ -161,7 +161,8 @@ impl<S: Services> ForgeApp<S> {
             .error_tracker(ToolErrorTracker::new(max_tool_failure_per_turn))
             .tool_definitions(tool_definitions)
             .models(models)
-            .hook(Arc::new(hook));
+            .hook(Arc::new(hook))
+            .tool_search(forge_config.tool_search);
 
         // Create and return the stream
         let stream = MpscStream::spawn(

crates/forge_app/src/dto/anthropic/request.rs

@@ -280,6 +280,10 @@ impl TryFrom<ContextMessage> for Message {
             ContextMessage::Tool(tool_result) => {
                 Message { role: Role::User, content: vec![tool_result.try_into()?] }
             }
+            ContextMessage::ToolSearchOutput(_) => {
+                // Tool search output is OpenAI Responses API specific - skip for Anthropic
+                Message { role: Role::User, content: vec![] }
+            }
             ContextMessage::Image(img) => {
                 Message { content: vec![Content::from(img)], role: Role::User }
             }

crates/forge_app/src/dto/anthropic/response.rs

@@ -394,6 +394,7 @@ impl TryFrom<ContentBlock> for ChatCompletionMessage {
                         serde_json::to_string(&input)?
                     },
                     thought_signature: None,
+                    namespace: None,
                 })
             }
             ContentBlock::InputJsonDelta { partial_json } => {
@@ -402,6 +403,7 @@ impl TryFrom<ContentBlock> for ChatCompletionMessage {
                     name: None,
                     arguments_part: partial_json,
                     thought_signature: None,
+                    namespace: None,
                 })
             }
         };

crates/forge_app/src/dto/anthropic/transforms/auth_system_message.rs

@@ -91,6 +91,7 @@ mod tests {
             stream: None,
             response_format: None,
             initiator: None,
+            tool_search: None,
         };
 
         Request::try_from(context).unwrap()

crates/forge_app/src/dto/anthropic/transforms/set_cache.rs

@@ -102,6 +102,7 @@ mod tests {
             stream: None,
             response_format: None,
             initiator: None,
+            tool_search: None,
         };
 
         let request = Request::try_from(context).expect("Failed to convert context to request");
@@ -241,6 +242,7 @@ mod tests {
             stream: None,
             response_format: None,
             initiator: None,
+            tool_search: None,
         };
 
         let request = Request::try_from(context).expect("Failed to convert context to request");

crates/forge_app/src/dto/google/request.rs

@@ -474,6 +474,10 @@ impl From<ContextMessage> for Content {
         match message {
             ContextMessage::Text(text_message) => Content::from(text_message),
             ContextMessage::Tool(tool_result) => Content::from(tool_result),
+            ContextMessage::ToolSearchOutput(_) => {
+                // Tool search output is OpenAI Responses API specific - skip for Google
+                Content { role: None, parts: vec![] }
+            }
             ContextMessage::Image(image) => Content::from(image),
         }
     }
@@ -576,6 +580,7 @@ mod tests {
                 r#"{"file_path":"test.rs","old_string":"foo","new_string":"bar"}"#,
             ),
             thought_signature: None,
+            namespace: None,
         };
 
         // Convert to Google Part
@@ -615,12 +620,14 @@ mod tests {
                 call_id: None,
                 arguments: ToolCallArguments::from_json(r#"{"path":"file1.rs"}"#),
                 thought_signature: None,
+                namespace: None,
             },
             ToolCallFull {
                 name: ToolName::new("remove"),
                 call_id: None,
                 arguments: ToolCallArguments::from_json(r#"{"path":"file2.rs"}"#),
                 thought_signature: None,
+                namespace: None,
             },
         ];
 

crates/forge_app/src/dto/google/response.rs

@@ -431,6 +431,7 @@ impl TryFrom<Part> for ChatCompletionMessage {
                         name: Some(ToolName::new(function_call.name)),
                         arguments_part: serde_json::to_string(&function_call.args)?,
                         thought_signature,
+                        namespace: None,
                     },
                 ),
             ),

crates/forge_app/src/dto/openai/request.rs

@@ -504,6 +504,21 @@ impl From<ContextMessage> for Message {
                     extra_content: None,
                 }
             }
+            ContextMessage::ToolSearchOutput(_) => {
+                // Tool search output is OpenAI Responses API specific - skip for OpenAI
+                Message {
+                    role: Role::User,
+                    content: None,
+                    name: None,
+                    tool_call_id: None,
+                    tool_calls: None,
+                    reasoning_details: None,
+                    reasoning_text: None,
+                    reasoning_opaque: None,
+                    reasoning_content: None,
+                    extra_content: None,
+                }
+            }
         }
     }
 }
@@ -741,6 +756,7 @@ mod tests {
             name: ToolName::new("test_tool"),
             arguments: serde_json::json!({"key": "value"}).into(),
             thought_signature: None,
+            namespace: None,
         };
 
         let assistant_message = ContextMessage::Text(

crates/forge_app/src/dto/openai/response.rs

@@ -372,6 +372,7 @@ impl TryFrom<Response> for ChatCompletionMessage {
                                             &tool_call.function.arguments,
                                         )?,
                                         thought_signature,
+                                        namespace: None,
                                     });
                                 }
                             }
@@ -433,6 +434,7 @@ impl TryFrom<Response> for ChatCompletionMessage {
                                         name: tool_call.function.name.clone(),
                                         arguments_part: tool_call.function.arguments.clone(),
                                         thought_signature,
+                                        namespace: None,
                                     });
                                 }
                             }

crates/forge_app/src/dto/openai/transformers/drop_tool_call.rs

@@ -47,6 +47,7 @@ mod tests {
             name: ToolName::new("test_tool"),
             arguments: serde_json::json!({"key": "value"}).into(),
             thought_signature: None,
+            namespace: None,
         };
 
         let tool_result = ToolResult::new(ToolName::new("test_tool"))
@@ -72,6 +73,7 @@ mod tests {
             stream: None,
             response_format: None,
             initiator: None,
+            tool_search: None,
         };
 
         let request = Request::from(context);

crates/forge_app/src/dto/openai/transformers/set_cache.rs

@@ -88,6 +88,7 @@ mod tests {
             stream: None,
             response_format: None,
             initiator: None,
+            tool_search: None,
         };
 
         let request = Request::from(context);

crates/forge_app/src/hooks/doom_loop.rs

@@ -269,6 +269,7 @@ mod tests {
             reasoning_details: None,
             droppable: false,
             phase: None,
+            response_items: None,
         }
     }
 
@@ -405,6 +406,7 @@ mod tests {
             reasoning_details: None,
             droppable: false,
             phase: None,
+            response_items: None,
         };
 
         let user_msg = TextMessage {
@@ -417,6 +419,7 @@ mod tests {
             reasoning_details: None,
             droppable: false,
             phase: None,
+            response_items: None,
         };
 
         let assistant_msg_2 = TextMessage {
@@ -429,6 +432,7 @@ mod tests {
             reasoning_details: None,
             droppable: false,
             phase: None,
+            response_items: None,
         };
 
         let messages = [

crates/forge_app/src/hooks/tracing.rs

@@ -205,6 +205,8 @@ mod tests {
             usage: Default::default(),
             finish_reason: None,
             phase: None,
+            tool_search_output: None,
+            response_items: None,
         };
         let event = EventData::new(test_agent(), test_model_id(), ResponsePayload::new(message));
 
@@ -221,6 +223,7 @@ mod tests {
             call_id: Some(ToolCallId::new("test-id")),
             arguments: serde_json::json!({"key": "value"}).into(),
             thought_signature: None,
+            namespace: None,
         };
         let result = ToolResult::new(ToolName::from("test-tool"))
             .call_id(ToolCallId::new("test-id"))

crates/forge_app/src/orch.rs

@@ -25,6 +25,10 @@ pub struct Orchestrator<S> {
     agent: Agent,
     error_tracker: ToolErrorTracker,
     hook: Arc<Hook>,
+    /// Whether tool_search API is enabled for deferred tool loading.
+    /// When `true`, MCP tools are sent with `defer_loading: true` and a
+    /// `tool_search` tool is injected so the model can discover them on demand.
+    tool_search: Option<bool>,
 }
 
 impl<S: AgentService> Orchestrator<S> {
@@ -44,6 +48,7 @@ impl<S: AgentService> Orchestrator<S> {
             models: Default::default(),
             error_tracker: Default::default(),
             hook: Arc::new(Hook::default()),
+            tool_search: Default::default(),
         }
     }
 
@@ -153,10 +158,13 @@ impl<S: AgentService> Orchestrator<S> {
     async fn execute_chat_turn(
         &self,
         model_id: &ModelId,
-        context: Context,
+        mut context: Context,
         reasoning_supported: bool,
     ) -> anyhow::Result<ChatCompletionMessageFull> {
         let tool_supported = self.is_tool_supported()?;
+        // Propagate tool_search config to the context so the repository layer
+        // can decide whether to apply deferred tool loading.
+        context.tool_search = self.tool_search;
         let mut transformers = DefaultTransformation::default()
             .pipe(SortTools::new(self.agent.tool_order()))
             .pipe(NormalizeToolCallArguments::new())
@@ -268,8 +276,11 @@ impl<S: AgentService> Orchestrator<S> {
 
             // Turn is completed, if finish_reason is 'stop'. Gemini models return stop as
             // finish reason with tool calls.
-            is_complete =
-                message.finish_reason == Some(FinishReason::Stop) && message.tool_calls.is_empty();
+            // When tool_search_output is present, the model discovered tools via search
+            // and needs another turn to actually use them — do NOT mark as complete.
+            is_complete = message.finish_reason == Some(FinishReason::Stop)
+                && message.tool_calls.is_empty()
+                && message.tool_search_output.is_none();
 
             // Should yield if a tool is asking for a follow-up
             should_yield = is_complete
@@ -314,6 +325,8 @@ impl<S: AgentService> Orchestrator<S> {
                 message.usage,
                 tool_call_records,
                 message.phase,
+                message.tool_search_output.clone(),
+                message.response_items.clone(),
             );
 
             if self.error_tracker.limit_reached() {

crates/forge_app/src/user_prompt.rs

@@ -77,6 +77,7 @@ impl<S: AttachmentService> UserPromptGenerator<S> {
                 model: Some(self.agent.model.clone()),
                 droppable: true, // Droppable so it can be removed during context compression
                 phase: None,
+                response_items: None,
             };
             context = context.add_message(ContextMessage::Text(todo_message));
         }
@@ -123,6 +124,7 @@ impl<S: AttachmentService> UserPromptGenerator<S> {
                 model: Some(self.agent.model.clone()),
                 droppable: true, // Piped input is droppable
                 phase: None,
+                response_items: None,
             };
             context = context.add_message(ContextMessage::Text(piped_message));
         }
@@ -200,6 +202,7 @@ impl<S: AttachmentService> UserPromptGenerator<S> {
                 model: Some(self.agent.model.clone()),
                 droppable: false,
                 phase: None,
+                response_items: None,
             };
             context = context.add_message(ContextMessage::Text(message));
         }

crates/forge_config/.forge.toml

@@ -23,6 +23,7 @@ restricted = false
 sem_search_top_k = 10
 services_url = "https://api.forgecode.dev/"
 tool_supported = true
+tool_search = false
 tool_timeout_secs = 300
 top_k = 30
 top_p = 0.8

crates/forge_config/src/config.rs

@@ -253,6 +253,14 @@ pub struct ForgeConfig {
     #[serde(default)]
     pub tool_supported: bool,
 
+    /// Whether server-side tool search is enabled for models that support
+    /// deferred tool loading (e.g. GPT-5.4). When enabled, MCP tools are
+    /// sent with `defer_loading: true` and a `tool_search` tool is injected
+    /// so the API can discover them on demand. Defaults to `false`; set to
+    /// `true` to enable.
+    #[serde(default)]
+    pub tool_search: bool,
+
     /// Reasoning configuration applied to all agents; controls effort level,
     /// token budget, and visibility of the model's thinking process.
     #[serde(default, skip_serializing_if = "Option::is_none")]

crates/forge_domain/src/compact/summary.rs

@@ -274,6 +274,9 @@ impl From<&Context> for ContextSummary {
                     }
                 }
                 ContextMessage::Image(_) => {}
+                ContextMessage::ToolSearchOutput(_) => {
+                    // Tool search output is not included in the summary
+                }
             }
         }
 
@@ -895,6 +898,7 @@ mod tests {
             call_id: Some(ToolCallId::new("call_1")),
             arguments: ToolCallArguments::from_json(r#"{"title": "Bug report"}"#),
             thought_signature: None,
+            namespace: None,
         };
 
         let actual = extract_tool_info(&fixture, &[]);
@@ -987,6 +991,7 @@ mod tests {
                     r#"{"path": "/test", "pattern": "pattern"}"#,
                 ),
                 thought_signature: None,
+                namespace: None,
             }],
         )]);
 
@@ -1464,6 +1469,7 @@ mod tests {
                     r#"{"title": "Bug report", "body": "Description"}"#,
                 ),
                 thought_signature: None,
+                namespace: None,
             }],
         )]);
 
@@ -1493,6 +1499,7 @@ mod tests {
                     call_id: Some(ToolCallId::new("call_1")),
                     arguments: ToolCallArguments::from_json(r#"{"title": "Bug"}"#),
                     thought_signature: None,
+                    namespace: None,
                 }],
             ),
             tool_result("mcp_github_create_issue", "call_1", false),
@@ -1523,6 +1530,7 @@ mod tests {
                     call_id: Some(ToolCallId::new("call_1")),
                     arguments: ToolCallArguments::from_json(r#"{"title": "Bug"}"#),
                     thought_signature: None,
+                    namespace: None,
                 },
                 ToolCallFull {
                     name: ToolName::new("mcp_slack_post_message"),
@@ -1531,6 +1539,7 @@ mod tests {
                         r##"{"channel": "#dev", "text": "Hello"}"##,
                     ),
                     thought_signature: None,
+                    namespace: None,
                 },
             ],
         )]);
@@ -1566,6 +1575,7 @@ mod tests {
                     call_id: Some(ToolCallId::new("call_2")),
                     arguments: ToolCallArguments::from_json(r#"{"title": "Bug"}"#),
                     thought_signature: None,
+                    namespace: None,
                 },
                 ToolCatalog::tool_call_write("/test/output.txt", "result").call_id("call_3"),
             ],