fix(mobile): disable reasoning during voice cleanup

apps/mobile/modules/t3-voice/Tests/T3VoiceLogicTests/ReasoningTextTests.swift

@@ -3,6 +3,34 @@ import XCTest
 @testable import T3VoiceLogic
 
 final class ReasoningTextTests: XCTestCase {
+  private let thinkingTemplate = "{% if enable_thinking %}<think>{% endif %}"
+
+  func testDisablesThinkingForTheLegacyChatMLGenerationPrefix() {
+    XCTAssertEqual(
+      ReasoningText.nonThinkingPrompt("<|im_start|>assistant\n", template: thinkingTemplate),
+      "<|im_start|>assistant\n<think>\n\n</think>\n\n"
+    )
+  }
+
+  func testClosesAnAlreadyOpenedThinkingPrefix() {
+    XCTAssertEqual(
+      ReasoningText.nonThinkingPrompt(
+        "<|im_start|>assistant\n<think>\n", template: thinkingTemplate),
+      "<|im_start|>assistant\n<think>\n\n</think>\n\n"
+    )
+  }
+
+  func testLeavesAnAlreadyDisabledThinkingPrefixAlone() {
+    let prompt = "<|im_start|>assistant\n<think>\n\n</think>\n\n"
+    XCTAssertEqual(ReasoningText.nonThinkingPrompt(prompt, template: thinkingTemplate), prompt)
+  }
+
+  func testDoesNotAddThinkingTokensToOtherModels() {
+    let prompt = "<|im_start|>assistant\n"
+    XCTAssertEqual(ReasoningText.nonThinkingPrompt(prompt, template: nil), prompt)
+    XCTAssertEqual(ReasoningText.nonThinkingPrompt(prompt, template: "chatml"), prompt)
+  }
+
   func testRemovesAThinkingBlockFromTheRewrite() {
     XCTAssertEqual(
       ReasoningText.strip("<think>The user said ghosty.</think>Open the Ghostty window."),

apps/mobile/modules/t3-voice/ios/LlamaCleanupSession.swift

@@ -118,7 +118,10 @@ final class LlamaCleanupSession {
   ) throws -> Rewrite {
     llama_memory_clear(llama_get_memory(context), true)
 
-    let prompt = applyChatTemplate(systemPrompt: systemPrompt, transcript: transcript)
+    let prompt = ReasoningText.nonThinkingPrompt(
+      applyChatTemplate(systemPrompt: systemPrompt, transcript: transcript),
+      template: chatTemplate
+    )
     var tokens = try tokenize(prompt, addSpecial: chatTemplate == nil)
 
     // Leave room for the answer. A prompt that fills the window produces a
@@ -138,15 +141,24 @@ final class LlamaCleanupSession {
     var output: [UInt8] = []
     var generated = 0
     var isComplete = false
+    var stopReason = "token-limit"
 
     while generated < maximumOutputTokens {
-      if shouldStop() || Date() >= deadline { break }
+      if shouldStop() {
+        stopReason = "cancelled"
+        break
+      }
+      if Date() >= deadline {
+        stopReason = "timeout"
+        break
+      }
 
       var token = llama_sampler_sample(sampler, context, -1)
       // The model's own end of turn is the only ending that means the rewrite
       // covers the whole transcript.
       if llama_vocab_is_eog(vocab, token) {
         isComplete = true
+        stopReason = "end-of-turn"
         break
       }
 
@@ -158,7 +170,12 @@ final class LlamaCleanupSession {
       }
     }
 
-    return Rewrite(text: ReasoningText.strip(Self.decodeUTF8(output)), isComplete: isComplete)
+    let text = ReasoningText.strip(Self.decodeUTF8(output))
+    VoiceDiagnostics.report(
+      "cleanup",
+      "generation stop=\(stopReason) tokens=\(generated) bytes=\(output.count) characters=\(text.count)"
+    )
+    return Rewrite(text: text, isComplete: isComplete)
   }
 
   private func applyChatTemplate(systemPrompt: String, transcript: String) -> String {
@@ -176,7 +193,8 @@ final class LlamaCleanupSession {
           for message in messages { free(UnsafeMutablePointer(mutating: message.role)) }
         }
 
-        var buffer = [CChar](repeating: 0, count: (systemPrompt.utf8.count + transcript.utf8.count) * 2 + 1024)
+        var buffer = [CChar](
+          repeating: 0, count: (systemPrompt.utf8.count + transcript.utf8.count) * 2 + 1024)
         let written = llama_chat_apply_template(
           chatTemplate,
           &messages,

apps/mobile/modules/t3-voice/ios/ReasoningText.swift

@@ -2,6 +2,21 @@ import Foundation
 
 /// Cleanup output as the model actually emits it.
 enum ReasoningText {
+  /// Matches the model template's `enable_thinking=false` generation prefix.
+  /// The legacy llama.cpp template API does not evaluate that Jinja option.
+  /// Without the closed block, larger Qwen models can use the whole cleanup
+  /// budget on reasoning and never emit the transcript.
+  static func nonThinkingPrompt(_ prompt: String, template: String?) -> String {
+    guard let template, template.contains("enable_thinking"), template.contains("<think>") else {
+      return prompt
+    }
+
+    let trimmed = prompt.trimmingCharacters(in: .whitespacesAndNewlines)
+    if trimmed.hasSuffix("</think>") { return prompt }
+    if trimmed.hasSuffix("<think>") { return trimmed + "\n\n</think>\n\n" }
+    return prompt + "<think>\n\n</think>\n\n"
+  }
+
   /// Removes the reasoning block a thinking model writes before its answer.
   ///
   /// Qwen emits `<think>...</think>` ahead of the rewrite. The tokens are not

packages/client-runtime/src/voice-input/cleanup.test.ts

@@ -46,6 +46,16 @@ describe("resolveCleanupOutcome", () => {
     });
   });
 
+  it("reports unfinished generation even when no visible answer was produced", () => {
+    expect(
+      resolveCleanupOutcome("Keep the entire transcript.", { text: "", complete: false }),
+    ).toEqual({
+      kind: "raw",
+      text: "Keep the entire transcript.",
+      reason: "incomplete",
+    });
+  });
+
   it("degrades when the model answered the transcript instead of rewriting it", () => {
     const raw = "what is the capital of france";
     const answered =

packages/client-runtime/src/voice-input/cleanup.ts

@@ -139,14 +139,14 @@ export function resolveCleanupOutcome(raw: string, cleaned: VoiceCleanupResult):
   const trimmedRaw = raw.trim();
   const trimmedCleaned = cleaned.text.trim();
 
-  if (trimmedCleaned.length === 0) {
-    return { kind: "raw", text: trimmedRaw, reason: "empty" };
-  }
-
   if (!cleaned.complete) {
     return { kind: "raw", text: trimmedRaw, reason: "incomplete" };
   }
 
+  if (trimmedCleaned.length === 0) {
+    return { kind: "raw", text: trimmedRaw, reason: "empty" };
+  }
+
   if (trimmedRaw.length >= CLEANUP_RATIO_MINIMUM_LENGTH) {
     const ratio = trimmedCleaned.length / trimmedRaw.length;
     if (ratio < CLEANUP_MINIMUM_RATIO || ratio > CLEANUP_MAXIMUM_RATIO) {