Skip to content

Fastest and only macOS Dictation app with on-device STT and custom trained AI enhancement model. Windows, iOS testflight! Android, Linux waitlist open! A local Wispr Flow alternative. Follow https://x.com/fluidvoiceapp for more

fix(ai): split streamed reasoning when only </think> arrives (#1001)

WissssleyL14f8457161e5
diff --git a/Sources/Fluid/Services/ThinkingParsers.swift b/Sources/Fluid/Services/ThinkingParsers.swift
index 625fbd40..cad3ac31 100644
--- a/Sources/Fluid/Services/ThinkingParsers.swift
+++ b/Sources/Fluid/Services/ThinkingParsers.swift
@@ -113,6 +113,8 @@ nonisolated enum ThinkingParserFactory {
/// Standard parser for models that use `<think>...</think>` or `<thinking>...</thinking>` tags.
/// Used by: DeepSeek, Qwen, Claude (when thinking enabled), most open models
nonisolated struct StandardThinkingParser: ThinkingParser {
+ private var sawOpeningTag = false
+
mutating func processChunk(
_ chunk: String,
currentState: ThinkingParserState,
@@ -136,6 +138,7 @@ nonisolated struct StandardThinkingParser: ThinkingParser {
// Remove processed part including open tag
tagBuffer = String(tagBuffer[openRange.upperBound...])
newState = .inThinking
+ self.sawOpeningTag = true
}
}
@@ -180,9 +183,20 @@ nonisolated struct StandardThinkingParser: ThinkingParser {
}
func finalize(thinkingBuffer: [String], contentBuffer: [String], finalState: ThinkingParserState) -> (thinking: String, content: String) {
- let thinking = thinkingBuffer.joined()
+ var thinking = thinkingBuffer.joined()
var content = contentBuffer.joined()
+ // Chat templates that open the think block in the prompt (e.g. Qwen3 Thinking-2507)
+ // stream "reasoning</think>answer" with no opening tag. Split on that close the same
+ // way stripThinkingTags does for non-streaming responses.
+ let firstClose = ["</think>", "</thinking>"]
+ .compactMap { content.range(of: $0) }
+ .min { $0.lowerBound < $1.lowerBound }
+ if !self.sawOpeningTag, let closeRange = firstClose {
+ thinking += content[..<closeRange.lowerBound]
+ content = String(content[closeRange.upperBound...])
+ }
+
// Strip any remaining stray tags
content = content.replacingOccurrences(of: "</think>", with: "")
content = content.replacingOccurrences(of: "</thinking>", with: "")
diff --git a/Tests/FluidDictationIntegrationTests/LLMClientRequestBodyTests.swift b/Tests/FluidDictationIntegrationTests/LLMClientRequestBodyTests.swift
index e668313d..9224a33d 100644
--- a/Tests/FluidDictationIntegrationTests/LLMClientRequestBodyTests.swift
+++ b/Tests/FluidDictationIntegrationTests/LLMClientRequestBodyTests.swift
@@ -545,6 +545,44 @@ final class LLMClientStreamingTests: XCTestCase {
XCTAssertEqual(response.toolCalls.first?.getString("command"), "pwd")
}
+ /// Qwen3 Thinking-2507 templates open the think block in the prompt, so the stream
+ /// only carries the closing tag.
+ func testTagParserSplitsReasoningWhenOnlyCloseTagStreams() async throws {
+ let client = self.makeClient()
+ var config = LLMClient.Config(
+ messages: [["role": "user", "content": "Clean this up"]],
+ model: "qwen3-30b-a3b-thinking-2507",
+ baseURL: "https://issue-445.test/orphan-close/v1",
+ apiKey: "",
+ streaming: true
+ )
+ config.maxRetries = 1
+ config.timeoutSeconds = 5
+
+ let response = try await client.call(config)
+
+ XCTAssertEqual(response.thinking, "Reasoning.")
+ XCTAssertEqual(response.content, "Ready.")
+ }
+
+ func testTagParserSplitsOnEarliestCloseTag() async throws {
+ let client = self.makeClient()
+ var config = LLMClient.Config(
+ messages: [["role": "user", "content": "Clean this up"]],
+ model: "qwen3-30b-a3b-thinking-2507",
+ baseURL: "https://issue-445.test/mixed-close/v1",
+ apiKey: "",
+ streaming: true
+ )
+ config.maxRetries = 1
+ config.timeoutSeconds = 5
+
+ let response = try await client.call(config)
+
+ XCTAssertEqual(response.thinking, "Reasoning.")
+ XCTAssertEqual(response.content, "Ready. Done.")
+ }
+
func testStreamingDecodeAndCallbacksStayOffMainThread() async throws {
let client = self.makeClient()
let probe = LLMCallbackThreadProbe()
@@ -615,6 +653,22 @@ private class Issue445StreamURLProtocol: URLProtocol {
"""#
+ private static let orphanCloseFixture = #"""
+ data: {"choices":[{"index":0,"delta":{"content":"Reasoning.</thi"}}]}
+
+ data: {"choices":[{"index":0,"delta":{"content":"nk>Ready."},"finish_reason":"stop"}]}
+
+ data: [DONE]
+
+ """#
+
+ private static let mixedCloseFixture = #"""
+ data: {"choices":[{"index":0,"delta":{"content":"Reasoning.</thinking>Ready.</think> Done."},"finish_reason":"stop"}]}
+
+ data: [DONE]
+
+ """#
+
override class func canInit(with request: URLRequest) -> Bool {
request.url?.host == "issue-445.test"
}
@@ -638,7 +692,15 @@ private class Issue445StreamURLProtocol: URLProtocol {
return
}
- let fixture = url.path.contains("tag-parser") ? Self.tagParserFixture : Self.separateReasoningFixture
+ let fixture = if url.path.contains("orphan-close") {
+ Self.orphanCloseFixture
+ } else if url.path.contains("mixed-close") {
+ Self.mixedCloseFixture
+ } else if url.path.contains("tag-parser") {
+ Self.tagParserFixture
+ } else {
+ Self.separateReasoningFixture
+ }
client?.urlProtocol(self, didReceive: response, cacheStoragePolicy: .notAllowed)
client?.urlProtocol(self, didLoad: Data(fixture.utf8))