diff --git a/Dayflow/Dayflow/Core/AI/OllamaProvider+Networking.swift b/Dayflow/Dayflow/Core/AI/OllamaProvider+Networking.swift
index 362af4f35..bcc2cfa80 100644
--- a/Dayflow/Dayflow/Core/AI/OllamaProvider+Networking.swift
+++ b/Dayflow/Dayflow/Core/AI/OllamaProvider+Networking.swift
@@ -6,11 +6,21 @@
 import Foundation
 
 extension OllamaProvider {
+  /// Default output cap kept low because long generations are slow on local hardware.
+  static let defaultMaxOutputTokens = 4000
+
+  /// Override for reasoning models that need more headroom (#246):
+  /// `defaults write teleportlabs.com.Dayflow llmLocalMaxOutputTokens -int 32000`
+  static var configuredMaxOutputTokens: Int {
+    let configured = UserDefaults.standard.integer(forKey: "llmLocalMaxOutputTokens")
+    return configured > 0 ? configured : defaultMaxOutputTokens
+  }
+
   struct ChatRequest: Codable {
     let model: String
     let messages: [ChatMessage]
     var temperature: Double = 0.7
-    var max_tokens: Int = 4000
+    var max_tokens: Int = OllamaProvider.configuredMaxOutputTokens
     var stream: Bool = false
   }
 
@@ -231,7 +241,7 @@ extension OllamaProvider {
   // Helper method for text-only requests
   func callTextAPI(
     _ prompt: String, operation: String, expectJSON: Bool = false, batchId: Int64? = nil,
-    maxRetries: Int = 3, maxTokens: Int = 4000
+    maxRetries: Int = 3, maxTokens: Int = OllamaProvider.configuredMaxOutputTokens
   ) async throws -> String {
     let systemPrompt =
       expectJSON
@@ -269,7 +279,8 @@ extension OllamaProvider {
 // MARK: - Text Generation
 
 extension OllamaProvider {
-  func generateText(prompt: String, maxTokens: Int = 4000) async throws
+  func generateText(prompt: String, maxTokens: Int = OllamaProvider.configuredMaxOutputTokens)
+    async throws
     -> (text: String, log: LLMCall)
   {
     let callStart = Date()
diff --git a/Dayflow/DayflowTests/OllamaProviderMaxTokensTests.swift b/Dayflow/DayflowTests/OllamaProviderMaxTokensTests.swift
new file mode 100644
index 000000000..9e729d37e
--- /dev/null
+++ b/Dayflow/DayflowTests/OllamaProviderMaxTokensTests.swift
@@ -0,0 +1,75 @@
+import XCTest
+
+@testable import Dayflow
+
+final class OllamaProviderMaxTokensTests: XCTestCase {
+  private let key = "llmLocalMaxOutputTokens"
+  private var savedValue: Any?
+
+  override func setUp() {
+    super.setUp()
+    savedValue = UserDefaults.standard.object(forKey: key)
+    UserDefaults.standard.removeObject(forKey: key)
+  }
+
+  override func tearDown() {
+    UserDefaults.standard.removeObject(forKey: key)
+    if let savedValue {
+      UserDefaults.standard.set(savedValue, forKey: key)
+    }
+    savedValue = nil
+    super.tearDown()
+  }
+
+  func testMaxOutputTokensDefaultsTo4000WhenUnset() {
+    XCTAssertEqual(OllamaProvider.configuredMaxOutputTokens, 4000)
+  }
+
+  func testMaxOutputTokensReadsUserDefaultsOverride() {
+    UserDefaults.standard.set(32000, forKey: key)
+    XCTAssertEqual(OllamaProvider.configuredMaxOutputTokens, 32000)
+  }
+
+  func testMaxOutputTokensFallsBackToDefaultForInvalidValues() {
+    UserDefaults.standard.set(0, forKey: key)
+    XCTAssertEqual(OllamaProvider.configuredMaxOutputTokens, 4000)
+
+    UserDefaults.standard.set(-100, forKey: key)
+    XCTAssertEqual(OllamaProvider.configuredMaxOutputTokens, 4000)
+  }
+
+  func testChatRequestUsesConfiguredMaxOutputTokens() {
+    UserDefaults.standard.set(16000, forKey: key)
+    let request = OllamaProvider.ChatRequest(model: "test-model", messages: [])
+    XCTAssertEqual(request.max_tokens, 16000)
+  }
+
+  // Boundary: exactly 1 is the smallest value the `configured > 0` guard
+  // treats as a real override (0 falls back, per the test above). Pins the
+  // guard's exclusive-zero semantics so a future `>= 0` regression is caught.
+  func testMaxOutputTokensAcceptsMinimumPositiveOverride() {
+    UserDefaults.standard.set(1, forKey: key)
+    XCTAssertEqual(OllamaProvider.configuredMaxOutputTokens, 1)
+  }
+
+  // Reasoning models (the #246 use case) can legitimately need very large
+  // caps. Confirm a large override passes through unclamped, both via the
+  // computed property and into the ChatRequest the network layer sends.
+  func testMaxOutputTokensPassesLargeReasoningModelOverrideThrough() {
+    UserDefaults.standard.set(131_072, forKey: key)
+    XCTAssertEqual(OllamaProvider.configuredMaxOutputTokens, 131_072)
+    let request = OllamaProvider.ChatRequest(model: "test-model", messages: [])
+    XCTAssertEqual(request.max_tokens, 131_072)
+  }
+
+  // The default constant and the fallback path must agree — guards against a
+  // future edit that changes one but not the other.
+  func testDefaultConstantMatchesUnsetFallback() {
+    XCTAssertEqual(OllamaProvider.defaultMaxOutputTokens, 4000)
+    UserDefaults.standard.removeObject(forKey: key)
+    XCTAssertEqual(
+      OllamaProvider.configuredMaxOutputTokens,
+      OllamaProvider.defaultMaxOutputTokens
+    )
+  }
+}