From f622912f2e1c30480082bd87f9a2dc81759136f2 Mon Sep 17 00:00:00 2001
From: AJ Isaacs <ajisaacs27@gmail.com>
Date: Sun, 22 Feb 2026 16:04:29 -0500
Subject: [PATCH] Improve: Support reasoning models in AI vision pipeline

Strip <think>...</think> blocks from reasoning model output (e.g.
Qwen3-VL-Thinking) and increase max_tokens from 4096 to 16384 to
accommodate thinking token overhead.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
---
 MoneyMap/Services/AIVisionClient.cs | 8 ++++++--
 1 file changed, 6 insertions(+), 2 deletions(-)

diff --git a/MoneyMap/Services/AIVisionClient.cs b/MoneyMap/Services/AIVisionClient.cs
index d289cfd..9290d7b 100644
--- a/MoneyMap/Services/AIVisionClient.cs
+++ b/MoneyMap/Services/AIVisionClient.cs
@@ -220,6 +220,10 @@ namespace MoneyMap.Services
         {
             var trimmed = content?.Trim() ?? "";
 
+            // Strip <think>...</think> blocks from reasoning models (e.g. Qwen3-VL-Thinking)
+            trimmed = System.Text.RegularExpressions.Regex.Replace(
+                trimmed, @"<think>[\s\S]*?</think>", "", System.Text.RegularExpressions.RegexOptions.IgnoreCase).Trim();
+
             // Strip markdown code fences
             if (trimmed.StartsWith("```json"))
             {
@@ -816,7 +820,7 @@ namespace MoneyMap.Services
                         }
                     }
                 },
-                max_tokens = 4096,
+                max_tokens = 16384,
                 temperature = 0.1
             };
 
@@ -898,7 +902,7 @@ namespace MoneyMap.Services
                     OpenAIToolUseHelper.BuildToolsArray(tools),
                     toolExecutor,
                     maxToolRounds,
-                    maxTokens: 4096,
+                    maxTokens: 16384,
                     _logger);
             }
             catch (TaskCanceledException ex) when (ex.InnerException is TimeoutException)