Stabilize ghost-text size, token-budget completion length, tabby model lineup, full-accept reset

.gitignore

@@ -27,6 +27,7 @@ venv/
 # Personal notes
 launch.txt
 posts.txt
+.writing/
 
 # Claude Code
 .claude/worktrees/

Cotabby.xcodeproj/project.pbxproj

@@ -30,6 +30,7 @@
 		E10000012F93000100DDD001 /* PromptContextSanitizerTests.swift in Sources */ = {isa = PBXBuildFile; fileRef = E10000112F93000100DDD011 /* PromptContextSanitizerTests.swift */; };
 		E10000022F93000100DDD002 /* PermissionAndContextModelTests.swift in Sources */ = {isa = PBXBuildFile; fileRef = E10000122F93000100DDD012 /* PermissionAndContextModelTests.swift */; };
 		E10000032F93000100DDD003 /* GhostSuggestionLayoutTests.swift in Sources */ = {isa = PBXBuildFile; fileRef = E10000132F93000100DDD013 /* GhostSuggestionLayoutTests.swift */; };
+		E10000992F93000100DDD099 /* GhostFontSizeStabilizerTests.swift in Sources */ = {isa = PBXBuildFile; fileRef = E10000A92F93000100DDD0A9 /* GhostFontSizeStabilizerTests.swift */; };
 		E10000042F93000100DDD004 /* BundledRuntimeLocatorTests.swift in Sources */ = {isa = PBXBuildFile; fileRef = E10000142F93000100DDD014 /* BundledRuntimeLocatorTests.swift */; };
 		E10000052F93000100DDD005 /* DisplayCoordinateConverterTests.swift in Sources */ = {isa = PBXBuildFile; fileRef = E10000152F93000100DDD015 /* DisplayCoordinateConverterTests.swift */; };
 		F10000012FA0000100EEE001 /* TextDirectionDetectorTests.swift in Sources */ = {isa = PBXBuildFile; fileRef = F10000112FA0000100EEE011 /* TextDirectionDetectorTests.swift */; };
@@ -69,6 +70,7 @@
 		E10000112F93000100DDD011 /* PromptContextSanitizerTests.swift */ = {isa = PBXFileReference; includeInIndex = 1; lastKnownFileType = sourcecode.swift; path = PromptContextSanitizerTests.swift; sourceTree = "<group>"; };
 		E10000122F93000100DDD012 /* PermissionAndContextModelTests.swift */ = {isa = PBXFileReference; includeInIndex = 1; lastKnownFileType = sourcecode.swift; path = PermissionAndContextModelTests.swift; sourceTree = "<group>"; };
 		E10000132F93000100DDD013 /* GhostSuggestionLayoutTests.swift */ = {isa = PBXFileReference; includeInIndex = 1; lastKnownFileType = sourcecode.swift; path = GhostSuggestionLayoutTests.swift; sourceTree = "<group>"; };
+		E10000A92F93000100DDD0A9 /* GhostFontSizeStabilizerTests.swift */ = {isa = PBXFileReference; includeInIndex = 1; lastKnownFileType = sourcecode.swift; path = GhostFontSizeStabilizerTests.swift; sourceTree = "<group>"; };
 		E10000142F93000100DDD014 /* BundledRuntimeLocatorTests.swift */ = {isa = PBXFileReference; includeInIndex = 1; lastKnownFileType = sourcecode.swift; path = BundledRuntimeLocatorTests.swift; sourceTree = "<group>"; };
 		E10000152F93000100DDD015 /* DisplayCoordinateConverterTests.swift */ = {isa = PBXFileReference; includeInIndex = 1; lastKnownFileType = sourcecode.swift; path = DisplayCoordinateConverterTests.swift; sourceTree = "<group>"; };
 		F10000112FA0000100EEE011 /* TextDirectionDetectorTests.swift */ = {isa = PBXFileReference; includeInIndex = 1; lastKnownFileType = sourcecode.swift; path = TextDirectionDetectorTests.swift; sourceTree = "<group>"; };
@@ -151,6 +153,7 @@
 				E10000112F93000100DDD011 /* PromptContextSanitizerTests.swift */,
 				E10000122F93000100DDD012 /* PermissionAndContextModelTests.swift */,
 				E10000132F93000100DDD013 /* GhostSuggestionLayoutTests.swift */,
+				E10000A92F93000100DDD0A9 /* GhostFontSizeStabilizerTests.swift */,
 				E10000142F93000100DDD014 /* BundledRuntimeLocatorTests.swift */,
 				E10000152F93000100DDD015 /* DisplayCoordinateConverterTests.swift */,
 				F10000112FA0000100EEE011 /* TextDirectionDetectorTests.swift */,
@@ -296,6 +299,7 @@
 				E10000012F93000100DDD001 /* PromptContextSanitizerTests.swift in Sources */,
 				E10000022F93000100DDD002 /* PermissionAndContextModelTests.swift in Sources */,
 				E10000032F93000100DDD003 /* GhostSuggestionLayoutTests.swift in Sources */,
+				E10000992F93000100DDD099 /* GhostFontSizeStabilizerTests.swift in Sources */,
 				E10000042F93000100DDD004 /* BundledRuntimeLocatorTests.swift in Sources */,
 				E10000052F93000100DDD005 /* DisplayCoordinateConverterTests.swift in Sources */,
 				F10000012FA0000100EEE001 /* TextDirectionDetectorTests.swift in Sources */,

Cotabby/App/Coordinators/SuggestionCoordinator+Acceptance.swift

@@ -108,9 +108,7 @@ extension SuggestionCoordinator {
             presentOverlay(
                 text: advancedSession.remainingText,
                 at: predictedCaret,
-                inputFrameRect: liveContext.inputFrameRect,
-                caretQuality: liveContext.caretQuality,
-                observedCharWidth: liveContext.observedCharWidth,
+                context: liveContext,
                 isRightToLeft: isRTL
             )
             schedulePostInsertionRefresh()
@@ -170,9 +168,7 @@ extension SuggestionCoordinator {
         presentOverlay(
             text: advancedSession.remainingText,
             at: session.baseContext.caretRect,
-            inputFrameRect: session.baseContext.inputFrameRect,
-            caretQuality: session.baseContext.caretQuality,
-            observedCharWidth: session.baseContext.observedCharWidth
+            context: session.baseContext
         )
         logStage(
             "typed-match-advanced",
@@ -314,17 +310,16 @@ extension SuggestionCoordinator {
     func presentOverlay(
         text: String,
         at caretRect: CGRect,
-        inputFrameRect: CGRect?,
-        caretQuality: CaretGeometryQuality,
-        observedCharWidth: CGFloat?,
+        context: FocusedInputContext,
         isRightToLeft: Bool = false
     ) {
         let geometry = SuggestionOverlayGeometry(
             caretRect: caretRect,
-            inputFrameRect: inputFrameRect,
-            caretQuality: caretQuality,
-            observedCharWidth: observedCharWidth,
-            isRightToLeft: isRightToLeft
+            inputFrameRect: context.inputFrameRect,
+            caretQuality: context.caretQuality,
+            observedCharWidth: context.observedCharWidth,
+            isRightToLeft: isRightToLeft,
+            focusChangeSequence: context.focusChangeSequence
         )
         if let message = overlayPresenter.present(
             text: text,

Cotabby/App/Coordinators/SuggestionCoordinator+Prediction.swift

@@ -224,9 +224,7 @@ extension SuggestionCoordinator {
         presentOverlay(
             text: session.remainingText,
             at: liveContext.caretRect,
-            inputFrameRect: liveContext.inputFrameRect,
-            caretQuality: liveContext.caretQuality,
-            observedCharWidth: liveContext.observedCharWidth,
+            context: liveContext,
             isRightToLeft: TextDirectionDetector.isRightToLeft(liveContext.precedingText)
         )
         logStage(
@@ -319,9 +317,7 @@ extension SuggestionCoordinator {
             presentOverlay(
                 text: reconciledSession.remainingText,
                 at: liveContext.caretRect,
-                inputFrameRect: liveContext.inputFrameRect,
-                caretQuality: liveContext.caretQuality,
-                observedCharWidth: liveContext.observedCharWidth,
+                context: liveContext,
                 isRightToLeft: TextDirectionDetector.isRightToLeft(liveContext.precedingText)
             )
             if let advancement {

Cotabby/Models/LlamaRuntimeModels.swift

@@ -86,11 +86,9 @@ enum RuntimeModelCatalog {
     static func displayName(for filename: String) -> String {
         switch filename {
         case "Qwen3-0.6B-Q4_K_M.gguf":
-            return "cotabby-swift-1"
-        case "Qwen3.5-0.8B-Q4_K_M.gguf":
-            return "cotabby-balanced-1"
+            return "tabby-fast-1"
         case "gemma-4-E2B-it-Q4_K_M.gguf":
-            return "cotabby-careful-1"
+            return "tabby-balanced-1"
         default:
             return filename
         }
@@ -115,17 +113,6 @@ enum RuntimeModelCatalog {
             expectedSizeBytes: 396_705_472,
             sha256: "ac2d97712095a558e31573f62f466a3f9d93990898b0ec79d7c974c1780d524a"
         ),
-        DownloadableRuntimeModel(
-            filename: "Qwen3.5-0.8B-Q4_K_M.gguf",
-            displayName: displayName(for: "Qwen3.5-0.8B-Q4_K_M.gguf"),
-            downloadURL: URL(
-                string:
-                    "https://huggingface.co/unsloth/Qwen3.5-0.8B-GGUF/resolve/main/Qwen3.5-0.8B-Q4_K_M.gguf?download=true"
-            )!,
-            approximateSizeInGigabytes: 0.5,
-            expectedSizeBytes: 532_517_120,
-            sha256: "bd258782e35f7f458f8aced1adc053e6e92e89bc735ba3be89d38a06121dc517"
-        ),
         DownloadableRuntimeModel(
             filename: "gemma-4-E2B-it-Q4_K_M.gguf",
             displayName: displayName(for: "gemma-4-E2B-it-Q4_K_M.gguf"),
@@ -153,7 +140,7 @@ struct LlamaRuntimeConfiguration: Equatable, Sendable {
     static let `default` = LlamaRuntimeConfiguration(
         runtimeDirectoryPath: nil,
         preferredModelNames: [
-            "Qwen3.5-0.8B-Q4_K_M.gguf",
+            "gemma-4-E2B-it-Q4_K_M.gguf",
             "Qwen3-0.6B-Q4_K_M.gguf"
         ],
         contextWindowTokens: 2048,

Cotabby/Models/SuggestionModels.swift

@@ -39,16 +39,17 @@ enum SuggestionWordCountPreset: String, CaseIterable, Equatable, Hashable, Senda
         }
     }
 
-    /// Token budget sized at ~1.5x the upper word bound. Tight enough to enforce the word cap
-    /// while leaving room for multi-token words (contractions, proper nouns, punctuation).
+    /// Token budget bumped 50% above the prior ~1.5x-upper-word-bound sizing, giving the
+    /// token-cap-only path (no in-prompt word range for the local model) room to land a clean
+    /// stopping point instead of hard-truncating mid-thought.
     var suggestedPredictionTokenBudget: Int {
         switch self {
         case .threeToSeven:
-            return 11
+            return 17
         case .sevenToTwelve:
-            return 18
+            return 27
         case .twelveToTwenty:
-            return 30
+            return 45
         }
     }
 }
@@ -352,6 +353,26 @@ struct SuggestionOverlayGeometry: Equatable, Sendable {
     /// When `true`, the text near the caret is Right-to-Left (Arabic, Hebrew, etc.) and the ghost
     /// text overlay should appear to the left of the caret instead of the right.
     let isRightToLeft: Bool
+    /// Identifies the focus session that produced this geometry. `OverlayController` keys its
+    /// per-session font-size stabilization on this value, so a field switch (or focus loss) starts
+    /// a fresh size baseline. Defaults to 0 for tests that do not exercise session-scoped behavior.
+    let focusChangeSequence: UInt64
+
+    init(
+        caretRect: CGRect,
+        inputFrameRect: CGRect?,
+        caretQuality: CaretGeometryQuality,
+        observedCharWidth: CGFloat?,
+        isRightToLeft: Bool,
+        focusChangeSequence: UInt64 = 0
+    ) {
+        self.caretRect = caretRect
+        self.inputFrameRect = inputFrameRect
+        self.caretQuality = caretQuality
+        self.observedCharWidth = observedCharWidth
+        self.isRightToLeft = isRightToLeft
+        self.focusChangeSequence = focusChangeSequence
+    }
 }
 
 /// The overlay is intentionally modeled as data so diagnostics can reason about visibility

Cotabby/Services/UI/OverlayController.swift

@@ -32,6 +32,11 @@ final class OverlayController: SuggestionOverlayControlling {
     /// instead of a full view rebuild + layout pass.
     private var hostingView: NSHostingView<GhostSuggestionView>?
 
+    /// Per-focus-session floor for caret-derived font size. Caret height flickers between the real
+    /// line height and the coarse field-height fallback from poll to poll; stabilizing keeps ghost
+    /// text from ballooning when the fallback wins. See `GhostFontSizeStabilizer`.
+    private var ghostFontStabilizer = GhostFontSizeStabilizer()
+
     init(suggestionSettings: SuggestionSettingsModel) {
         self.suggestionSettings = suggestionSettings
     }
@@ -66,8 +71,12 @@ final class OverlayController: SuggestionOverlayControlling {
             return
         }
 
+        let stabilizedCaretHeight = ghostFontStabilizer.stabilizedCaretHeight(
+            geometry.caretRect.height,
+            focusSessionKey: geometry.focusChangeSequence
+        )
         let fontSize = resolvedGhostFontSize(
-            for: geometry.caretRect,
+            forCaretHeight: stabilizedCaretHeight,
             caretQuality: geometry.caretQuality
         )
         let layout = GhostSuggestionLayout.make(
@@ -118,14 +127,15 @@ final class OverlayController: SuggestionOverlayControlling {
     /// Exact and derived caret rects usually reflect the real text line height, so they may scale
     /// up in larger editors. Estimated rects are much less trustworthy because some apps only
     /// expose the full field frame; the extra ceiling prevents one bad estimate from rendering
-    /// comically oversized ghost text.
+    /// comically oversized ghost text. `caretHeight` is already floored to the per-session minimum
+    /// by `ghostFontStabilizer`, so this only applies the static floor and quality ceilings.
     private func resolvedGhostFontSize(
-        for caretRect: CGRect,
+        forCaretHeight caretHeight: CGFloat,
         caretQuality: CaretGeometryQuality
     ) -> CGFloat {
         let proposedSize = max(
             Layout.minimumGhostFontSize,
-            caretRect.height * Layout.fontToLineHeightRatio
+            caretHeight * Layout.fontToLineHeightRatio
         )
         let qualityCap = caretQuality == .estimated
             ? Layout.maximumEstimatedGhostFontSize

Cotabby/Support/GhostFontSizeStabilizer.swift

@@ -0,0 +1,43 @@
+import CoreGraphics
+import Foundation
+
+/// Floors ghost-text size to the smallest caret line height observed during one focus session.
+///
+/// AX caret geometry is eventually consistent and app-specific. The same field can yield a tight
+/// line-height caret on one poll (zero-length `BoundsForRange`) and the full field-height `AXFrame`
+/// fallback on the next, when the precise branches happen to fail. Because `OverlayController`
+/// derives ghost font size from caret height, that fluctuation renders the suggestion comically
+/// oversized whenever the coarse fallback wins a poll.
+///
+/// Within a single focus session the real line height does not grow, so we treat the smallest
+/// height we have seen as the truth and clamp larger readings down to it. The baseline is keyed by
+/// `FocusTracker`'s `focusChangeSequence`, so switching fields — or leaving and re-entering the same
+/// field — starts a fresh measurement instead of inheriting a stale ceiling.
+///
+/// This intentionally biases toward the smaller reading: an over-tall fallback is the observed
+/// failure mode, and the downstream `minimumGhostFontSize` floor bounds how small a spurious low
+/// reading can make the text.
+struct GhostFontSizeStabilizer {
+    private var sessionKey: UInt64?
+    private var minCaretHeight: CGFloat?
+
+    /// Returns the caret height to derive font size from: the running per-session minimum.
+    ///
+    /// Non-positive heights (empty rects) pass through untouched so a transient bad poll can't pin
+    /// the session minimum to zero and force every later suggestion to the font-size floor.
+    mutating func stabilizedCaretHeight(_ caretHeight: CGFloat, focusSessionKey: UInt64) -> CGFloat {
+        guard caretHeight > 0 else {
+            return caretHeight
+        }
+
+        if sessionKey != focusSessionKey {
+            sessionKey = focusSessionKey
+            minCaretHeight = caretHeight
+            return caretHeight
+        }
+
+        let stabilized = min(caretHeight, minCaretHeight ?? caretHeight)
+        minCaretHeight = stabilized
+        return stabilized
+    }
+}

Cotabby/Support/LlamaPromptRenderer.swift

@@ -77,7 +77,11 @@ enum LlamaPromptRenderer {
         if let languageInstruction, !languageInstruction.isEmpty {
             sections.append("- \(languageInstruction)")
         }
-        sections.append("- \(completionLengthInstruction)")
+        // Experiment: the explicit word-range line (`completionLengthInstruction`) is intentionally
+        // omitted from the local-model prompt so length is governed purely by the token budget
+        // (`SuggestionWordCountPreset.suggestedPredictionTokenBudget`). The parameter stays wired so
+        // re-enabling the in-prompt cue is a one-line change. Apple Intelligence still gets the cue.
+        _ = completionLengthInstruction
         sections.append("- The next line must begin directly with the continuation text.")
         sections.append("Text before caret:")
         sections.append(prefixText)

Cotabby/UI/SettingsView.swift

@@ -307,6 +307,16 @@ struct SettingsView: View {
                         }
                     }
 
+                    if suggestionSettings.fullAcceptanceKeyCode != SuggestionSettingsModel.defaultFullAcceptanceKeyCode {
+                        Button("Reset") {
+                            suggestionSettings.setFullAcceptanceKey(
+                                keyCode: SuggestionSettingsModel.defaultFullAcceptanceKeyCode,
+                                label: SuggestionSettingsModel.defaultFullAcceptanceKeyLabel
+                            )
+                            isRecordingFullAcceptKeybind = false
+                        }
+                    }
+
                     if suggestionSettings.fullAcceptanceKeyCode != SuggestionSettingsModel.disabledKeyCode {
                         Button("Clear") {
                             suggestionSettings.clearFullAcceptanceKey()

Cotabby/UI/WelcomeView.swift

@@ -322,7 +322,10 @@ extension WelcomeView {
                         suggestionSettings.setFullAcceptanceKey(keyCode: keyCode, label: label)
                     },
                     onReset: suggestionSettings.fullAcceptanceKeyCode != SuggestionSettingsModel.defaultFullAcceptanceKeyCode ? {
-                        suggestionSettings.clearFullAcceptanceKey()
+                        suggestionSettings.setFullAcceptanceKey(
+                            keyCode: SuggestionSettingsModel.defaultFullAcceptanceKeyCode,
+                            label: SuggestionSettingsModel.defaultFullAcceptanceKeyLabel
+                        )
                     } : nil
                 )
             }

CotabbyTests/CustomRulesTests.swift

@@ -102,7 +102,9 @@ final class CustomRulesTests: XCTestCase {
         XCTAssertTrue(instruction?.contains("Spanish") == true)
     }
 
-    func test_llamaRenderer_includesLanguageInstructionBeforeLengthCue() {
+    func test_llamaRenderer_includesLanguageInstructionInFinalBlock() {
+        // The length cue is no longer rendered (token-budget-only experiment), so this guards that
+        // the language directive still lands in the late, high-attention final-instruction block.
         let prompt = LlamaPromptRenderer.prompt(
             prefixText: "Hola",
             applicationName: "Notes",
@@ -111,12 +113,14 @@ final class CustomRulesTests: XCTestCase {
             languageInstruction: SuggestionLanguage.spanish.promptInstruction
         )
 
-        guard let langRange = prompt.range(of: "Spanish"),
-              let lenRange = prompt.range(of: "UNIQUE_LENGTH_CUE") else {
-            XCTFail("Expected language directive and length cue in the prompt")
+        XCTAssertFalse(prompt.contains("UNIQUE_LENGTH_CUE"))
+
+        guard let finalRange = prompt.range(of: "Final instruction:"),
+              let langRange = prompt.range(of: "Spanish") else {
+            XCTFail("Expected final instruction header and language directive in the prompt")
             return
         }
-        XCTAssertLessThan(langRange.lowerBound, lenRange.lowerBound)
+        XCTAssertLessThan(finalRange.lowerBound, langRange.lowerBound)
     }
 
     func test_foundationModelInstructions_includeLanguageOverride() {