fix(dictation): stop leaking raw <think> content when stripped output is empty

Fluid.xcodeproj/project.pbxproj

@@ -12,6 +12,7 @@
 		7C3697892ED70F9C005874CE /* DynamicNotchKit in Frameworks */ = {isa = PBXBuildFile; productRef = 7C3697882ED70F9C005874CE /* DynamicNotchKit */; };
 		7C5AF14B2F15041600DE21B0 /* MediaRemoteAdapter in Frameworks */ = {isa = PBXBuildFile; productRef = 7C5AF14A2F15041600DE21B0 /* MediaRemoteAdapter */; };
 		7C91B0012F42AA0100C0DEF0 /* HotkeyShortcutTests.swift in Sources */ = {isa = PBXBuildFile; fileRef = 7C91B0022F42AA0100C0DEF0 /* HotkeyShortcutTests.swift */; };
+		9E1636273C4735E04F0940B0 /* LLMClientThinkingTagTests.swift in Sources */ = {isa = PBXBuildFile; fileRef = 18FD426ADEB23850ABB2FFEA /* LLMClientThinkingTagTests.swift */; };
 		7CDB0A2D2F3C4D5600FB7CAD /* DictationE2ETests.swift in Sources */ = {isa = PBXBuildFile; fileRef = 7CDB0A292F3C4D5600FB7CAD /* DictationE2ETests.swift */; };
 		7CDB0A2E2F3C4D5600FB7CAD /* AudioFixtureLoader.swift in Sources */ = {isa = PBXBuildFile; fileRef = 7CDB0A2A2F3C4D5600FB7CAD /* AudioFixtureLoader.swift */; };
 		7CDB0A2F2F3C4D5600FB7CAD /* dictation_fixture.wav in Resources */ = {isa = PBXBuildFile; fileRef = 7CDB0A2B2F3C4D5600FB7CAD /* dictation_fixture.wav */; };
@@ -33,6 +34,7 @@
 		7C078D8F2E3B339200FB7CAC /* FluidVoice Debug.app */ = {isa = PBXFileReference; explicitFileType = wrapper.application; includeInIndex = 0; path = "FluidVoice Debug.app"; sourceTree = BUILT_PRODUCTS_DIR; };
 		7CDB0A202F3C4D5600FB7CAD /* FluidDictationIntegrationTests.xctest */ = {isa = PBXFileReference; explicitFileType = wrapper.cfbundle; includeInIndex = 0; path = FluidDictationIntegrationTests.xctest; sourceTree = BUILT_PRODUCTS_DIR; };
 		7C91B0022F42AA0100C0DEF0 /* HotkeyShortcutTests.swift */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.swift; path = HotkeyShortcutTests.swift; sourceTree = "<group>"; };
+		18FD426ADEB23850ABB2FFEA /* LLMClientThinkingTagTests.swift */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.swift; path = LLMClientThinkingTagTests.swift; sourceTree = "<group>"; };
 		7CDB0A292F3C4D5600FB7CAD /* DictationE2ETests.swift */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.swift; path = DictationE2ETests.swift; sourceTree = "<group>"; };
 		7CDB0A2A2F3C4D5600FB7CAD /* AudioFixtureLoader.swift */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.swift; path = AudioFixtureLoader.swift; sourceTree = "<group>"; };
 		7CDB0A2B2F3C4D5600FB7CAD /* dictation_fixture.wav */ = {isa = PBXFileReference; lastKnownFileType = audio.wav; path = dictation_fixture.wav; sourceTree = "<group>"; };
@@ -104,6 +106,7 @@
 				7CDB0A272F3C4D5600FB7CAD /* Resources */,
 				7CDB0A292F3C4D5600FB7CAD /* DictationE2ETests.swift */,
 				7C91B0022F42AA0100C0DEF0 /* HotkeyShortcutTests.swift */,
+				18FD426ADEB23850ABB2FFEA /* LLMClientThinkingTagTests.swift */,
 			);
 			path = FluidDictationIntegrationTests;
 			sourceTree = "<group>";
@@ -258,6 +261,7 @@
 				7CDB0A2E2F3C4D5600FB7CAD /* AudioFixtureLoader.swift in Sources */,
 				7CDB0A2D2F3C4D5600FB7CAD /* DictationE2ETests.swift in Sources */,
 				7C91B0012F42AA0100C0DEF0 /* HotkeyShortcutTests.swift in Sources */,
+				9E1636273C4735E04F0940B0 /* LLMClientThinkingTagTests.swift in Sources */,
 			);
 			runOnlyForDeploymentPostprocessing = 0;
 		};

Sources/Fluid/Services/LLMClient.swift

@@ -795,7 +795,7 @@ final class LLMClient {
 
     // MARK: - Parse Non-Streaming Message
 
-    private func parseResponsesResponse(_ json: [String: Any]) throws -> Response {
+    func parseResponsesResponse(_ json: [String: Any]) throws -> Response {
         guard let output = json["output"] as? [[String: Any]] else {
             throw LLMError.invalidResponse
         }
@@ -838,12 +838,12 @@ final class LLMClient {
 
         return Response(
             thinking: thinking.isEmpty ? nil : thinking,
-            content: cleanedContent.isEmpty ? rawContent.trimmingCharacters(in: .whitespacesAndNewlines) : cleanedContent,
+            content: cleanedContent,
             toolCalls: parsedToolCalls
         )
     }
 
-    private func parseMessageResponse(_ message: [String: Any]) -> Response {
+    func parseMessageResponse(_ message: [String: Any]) -> Response {
         // Extract content
         let rawContent = message["content"] as? String ?? ""
 
@@ -878,7 +878,7 @@ final class LLMClient {
 
         return Response(
             thinking: finalThinking.isEmpty ? nil : finalThinking,
-            content: cleanedContent.isEmpty ? rawContent : cleanedContent,
+            content: cleanedContent,
             toolCalls: parsedToolCalls
         )
     }

Tests/FluidDictationIntegrationTests/LLMClientThinkingTagTests.swift

@@ -0,0 +1,65 @@
+@testable import FluidVoice_Debug
+import Foundation
+import XCTest
+
+@MainActor
+final class LLMClientThinkingTagTests: XCTestCase {
+    private let client = LLMClient.shared
+
+    // MARK: - Chat-completions message parser
+
+    func testMessageResponseDropsContentThatIsAllThinking() {
+        let message: [String: Any] = ["content": "<think>reasoning</think>"]
+
+        let response = self.client.parseMessageResponse(message)
+
+        XCTAssertTrue(response.content.isEmpty, "All-thinking replies must not leak raw <think> content downstream")
+        XCTAssertEqual(response.thinking, "reasoning")
+    }
+
+    func testMessageResponseKeepsContentAfterThinking() {
+        let message: [String: Any] = ["content": "<think>reasoning</think>Hello there"]
+
+        let response = self.client.parseMessageResponse(message)
+
+        XCTAssertEqual(response.content, "Hello there")
+        XCTAssertEqual(response.thinking, "reasoning")
+    }
+
+    func testMessageResponseLeavesPlainContentUnchanged() {
+        let message: [String: Any] = ["content": "Hello there"]
+
+        let response = self.client.parseMessageResponse(message)
+
+        XCTAssertEqual(response.content, "Hello there")
+        XCTAssertNil(response.thinking)
+    }
+
+    // MARK: - Responses API parser
+
+    func testResponsesResponseDropsContentThatIsAllThinking() throws {
+        let json = self.responsesPayload(text: "<think>reasoning</think>")
+
+        let response = try self.client.parseResponsesResponse(json)
+
+        XCTAssertTrue(response.content.isEmpty, "All-thinking replies must not leak raw <think> content downstream")
+        XCTAssertEqual(response.thinking, "reasoning")
+    }
+
+    func testResponsesResponseKeepsContentAfterThinking() throws {
+        let json = self.responsesPayload(text: "<think>reasoning</think>Hello there")
+
+        let response = try self.client.parseResponsesResponse(json)
+
+        XCTAssertEqual(response.content, "Hello there")
+        XCTAssertEqual(response.thinking, "reasoning")
+    }
+
+    // MARK: - Helpers
+
+    private func responsesPayload(text: String) -> [String: Any] {
+        let part: [String: Any] = ["type": "output_text", "text": text]
+        let message: [String: Any] = ["type": "message", "content": [part]]
+        return ["output": [message]]
+    }
+}