feat(tools,context): Phase 1 tool calling + Phase 2 context management

Phase 1 - Reliable Tool Calling:
- Add chat_generate command to Python daemon using tokenizer.apply_chat_template()
  with ChatML fallback for proper model-specific formatting
- Send structured JSON message arrays instead of flat prompt strings
- Classify 40+ tools into Core/Development/Media/Advanced tiers, only include
  4-7 tools in system prompt instead of all 40+
- New compact JSON tool format: <tool>{"name":"..","args":{..}}</tool>
  with few-shot examples for model guidance
- Tool approval flow: auto-approve read-only, show UI for write/execute
- Inline ToolApprovalView with approve/deny buttons (Return/Escape shortcuts)
- Restore maxTokens to 2048, maxResponseLength to 16000, maxResponseTokens to 4096

Phase 2 - Context Management:
- ContextBudget system allocates tokens based on model's actual context window
  (reads max_position_embeddings from config.json with name heuristic fallback)
- Word-based token estimation (~1.3 tok/word English, ~1.5 code)
- Budget-aware assembly: system(fixed) + recent(70%) + project(20%) + summary(10%)
- Rule-based compaction for dropped messages (instant, no LLM calls)
- Auto-include project file tree and recent files within budget

New files: ToolTier.swift, ToolApproval.swift, ToolApprovalView.swift, ContextBudget.swift
Modified: 12 files across daemon, services, models, tools, viewmodels, views
Version: 4.0.0 build 3

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

Author: kochj23

MLX Code.xcodeproj/project.pbxproj

@@ -13,5 +13,9 @@
 			<string>com.apple.dt.Xcode</string>
 		</array>
 	</dict>
+	<key>com.apple.security.application-groups</key>
+	<array>
+		<string>group.com.jkoch.mlxcode</string>
+	</array>
 </dict>
 </plist>

MLX Code/MLX_Code.entitlements

@@ -35,6 +35,9 @@ struct MLXModel: Identifiable, Codable, Equatable, Hashable {
     /// Model description
     var description: String?
 
+    /// Maximum context window size in tokens (detected from model config)
+    var contextWindowSize: Int?
+
     /// Initializes a new MLX model configuration
     /// - Parameters:
     ///   - id: Unique identifier (defaults to new UUID)
@@ -53,7 +56,8 @@ struct MLXModel: Identifiable, Codable, Equatable, Hashable {
         isDownloaded: Bool = false,
         sizeInBytes: Int64? = nil,
         huggingFaceId: String? = nil,
-        description: String? = nil
+        description: String? = nil,
+        contextWindowSize: Int? = nil
     ) {
         self.id = id
         self.name = name
@@ -63,6 +67,7 @@ struct MLXModel: Identifiable, Codable, Equatable, Hashable {
         self.sizeInBytes = sizeInBytes
         self.huggingFaceId = huggingFaceId
         self.description = description
+        self.contextWindowSize = contextWindowSize
     }
 }
 
@@ -96,11 +101,11 @@ struct ModelParameters: Codable, Equatable, Hashable {
     ///   - repetitionContextSize: Context size for repetition (default: 64, look back farther)
     init(
         temperature: Double = 0.7,
-        maxTokens: Int = 512,  // Reduced from 2048 to prevent infinite generation
+        maxTokens: Int = 2048,  // Restored: chat templates + RepetitionDetector prevent loops
         topP: Double = 0.9,
         topK: Int = 40,
-        repetitionPenalty: Double = 1.8,  // Increased from 1.2 to heavily penalize repetition
-        repetitionContextSize: Int = 256  // Increased from 64 to look back farther for repetition
+        repetitionPenalty: Double = 1.8,  // Keep high to penalize repetition
+        repetitionContextSize: Int = 256  // Keep wide lookback for repetition detection
     ) {
         self.temperature = temperature
         self.maxTokens = maxTokens
@@ -227,86 +232,95 @@ extension MLXModel {
             MLXModel(
                 name: "Qwen 2.5 7B ⭐ RECOMMENDED",
                 path: "\(modelBasePath)/qwen-2.5-7b",
-                parameters: ModelParameters(temperature: 0.7, maxTokens: 512, topP: 0.9, topK: 40, repetitionPenalty: 1.8, repetitionContextSize: 256),
+                parameters: ModelParameters(),
                 isDownloaded: false,
                 huggingFaceId: "mlx-community/Qwen2.5-7B-Instruct-4bit",
-                description: "Alibaba's Qwen 2.5 - Excellent quality, good instruction following. Best for coding. ~4GB"
+                description: "Alibaba's Qwen 2.5 - Excellent quality, good instruction following. Best for coding. ~4GB",
+                contextWindowSize: 32768
             ),
 
             MLXModel(
                 name: "Mistral 7B v0.3 ⭐ POPULAR",
                 path: "\(modelBasePath)/mistral-7b",
-                parameters: ModelParameters(temperature: 0.7, maxTokens: 512, topP: 0.9, topK: 40, repetitionPenalty: 1.8, repetitionContextSize: 256),
+                parameters: ModelParameters(),
                 isDownloaded: false,
                 huggingFaceId: "mlx-community/Mistral-7B-Instruct-v0.3-4bit",
-                description: "Mistral AI's 7B - Very popular, versatile, good for code. ~4GB"
+                description: "Mistral AI's 7B - Very popular, versatile, good for code. ~4GB",
+                contextWindowSize: 32768
             ),
 
             // Larger Models (Better Quality)
             MLXModel(
                 name: "Qwen 2.5 14B (Best Quality)",
                 path: "\(modelBasePath)/qwen-2.5-14b",
-                parameters: ModelParameters(temperature: 0.7, maxTokens: 512, topP: 0.9, topK: 40, repetitionPenalty: 1.8, repetitionContextSize: 256),
+                parameters: ModelParameters(),
                 isDownloaded: false,
                 huggingFaceId: "mlx-community/Qwen2.5-14B-Instruct-4bit",
-                description: "Qwen 2.5 14B - Best quality, slower but excellent results. ~8GB"
+                description: "Qwen 2.5 14B - Best quality, slower but excellent results. ~8GB",
+                contextWindowSize: 32768
             ),
 
             MLXModel(
                 name: "Llama 3.1 8B",
                 path: "\(modelBasePath)/llama-3.1-8b",
-                parameters: ModelParameters(temperature: 0.7, maxTokens: 512, topP: 0.9, topK: 40, repetitionPenalty: 1.8, repetitionContextSize: 256),
+                parameters: ModelParameters(),
                 isDownloaded: false,
                 huggingFaceId: "mlx-community/Meta-Llama-3.1-8B-Instruct-4bit",
-                description: "Meta's Llama 3.1 8B - Solid performance, good for general use. ~5GB"
+                description: "Meta's Llama 3.1 8B - Solid performance, good for general use. ~5GB",
+                contextWindowSize: 131072
             ),
 
             // Specialized Models
             MLXModel(
                 name: "DeepSeek Coder 6.7B",
                 path: "\(modelBasePath)/deepseek-coder",
-                parameters: ModelParameters(temperature: 0.7, maxTokens: 512, topP: 0.9, topK: 40, repetitionPenalty: 1.8, repetitionContextSize: 256),
+                parameters: ModelParameters(),
                 isDownloaded: false,
                 huggingFaceId: "mlx-community/deepseek-coder-6.7b-instruct",
-                description: "DeepSeek Coder - Specialized for code, excellent for programming tasks. ~4GB"
+                description: "DeepSeek Coder - Specialized for code, excellent for programming tasks. ~4GB",
+                contextWindowSize: 16384
             ),
 
             MLXModel(
                 name: "CodeLlama 7B",
                 path: "\(modelBasePath)/codellama-7b",
-                parameters: ModelParameters(temperature: 0.7, maxTokens: 512, topP: 0.9, topK: 40, repetitionPenalty: 1.8, repetitionContextSize: 256),
+                parameters: ModelParameters(),
                 isDownloaded: false,
                 huggingFaceId: "mlx-community/CodeLlama-7b-Instruct-hf-4bit-MLX",
-                description: "Meta's CodeLlama - Trained specifically for code generation. ~4GB"
+                description: "Meta's CodeLlama - Trained specifically for code generation. ~4GB",
+                contextWindowSize: 16384
             ),
 
             // Compact Models (Faster)
             MLXModel(
                 name: "Phi-3.5 Mini (Fast)",
                 path: "\(modelBasePath)/phi-3.5-mini",
-                parameters: ModelParameters(temperature: 0.7, maxTokens: 512, topP: 0.9, topK: 40, repetitionPenalty: 1.8, repetitionContextSize: 256),
+                parameters: ModelParameters(),
                 isDownloaded: false,
                 huggingFaceId: "mlx-community/Phi-3.5-mini-instruct-4bit",
-                description: "Microsoft's Phi-3.5 - Small but capable, very fast. ~2GB"
+                description: "Microsoft's Phi-3.5 - Small but capable, very fast. ~2GB",
+                contextWindowSize: 4096
             ),
 
             MLXModel(
                 name: "Gemma 2 9B",
                 path: "\(modelBasePath)/gemma-2-9b",
-                parameters: ModelParameters(temperature: 0.7, maxTokens: 512, topP: 0.9, topK: 40, repetitionPenalty: 1.8, repetitionContextSize: 256),
+                parameters: ModelParameters(),
                 isDownloaded: false,
                 huggingFaceId: "mlx-community/gemma-2-9b-it-4bit",
-                description: "Google's Gemma 2 9B - Excellent quality, good instruction following. ~5GB"
+                description: "Google's Gemma 2 9B - Excellent quality, good instruction following. ~5GB",
+                contextWindowSize: 8192
             ),
 
             // Keep original for backwards compatibility
             MLXModel(
                 name: "Llama 3.2 3B (Small - NOT RECOMMENDED)",
                 path: "\(modelBasePath)/llama-3.2-3b",
-                parameters: ModelParameters(temperature: 0.7, maxTokens: 512, topP: 0.9, topK: 40, repetitionPenalty: 1.8, repetitionContextSize: 256),
+                parameters: ModelParameters(),
                 isDownloaded: false,
                 huggingFaceId: "mlx-community/Llama-3.2-3B-Instruct-4bit",
-                description: "Meta's Llama 3.2 3B - Too small, causes loops. Use Qwen or Mistral instead. ~2GB"
+                description: "Meta's Llama 3.2 3B - Too small, causes loops. Use Qwen or Mistral instead. ~2GB",
+                contextWindowSize: 8192
             )
         ]
     }

MLX Code/Models/MLXModel.swift

@@ -0,0 +1,87 @@
+//
+//  ToolApproval.swift
+//  MLX Code
+//
+//  Tool approval types and policy for the agentic loop.
+//  Created on 2026-02-19.
+//
+
+import Foundation
+
+/// Represents a pending tool call awaiting user approval
+struct PendingToolCall: Identifiable {
+    let id = UUID()
+    let toolName: String
+    let parameters: [String: Any]
+    let rawJSON: String
+    var approved: Bool = false
+
+    /// Human-readable summary of the tool call
+    var summary: String {
+        if let path = parameters["path"] as? String {
+            return "\(toolName) → \(path)"
+        } else if let command = parameters["command"] as? String {
+            return "\(toolName) → \(command.prefix(60))"
+        } else if let pattern = parameters["pattern"] as? String {
+            return "\(toolName) → \(pattern)"
+        }
+        return toolName
+    }
+}
+
+/// Tool approval policy
+enum ToolApprovalPolicy: String, CaseIterable, Identifiable {
+    /// Always ask before executing any tool
+    case alwaysAsk = "always_ask"
+    /// Auto-approve read-only tools (file read, grep, glob, code navigation)
+    case autoApproveRead = "auto_approve_read"
+    /// Auto-approve all tools without asking
+    case autoApproveAll = "auto_approve_all"
+
+    var id: String { rawValue }
+
+    var displayName: String {
+        switch self {
+        case .alwaysAsk: return "Always Ask"
+        case .autoApproveRead: return "Auto-approve Read-only"
+        case .autoApproveAll: return "Auto-approve All"
+        }
+    }
+
+    var description: String {
+        switch self {
+        case .alwaysAsk:
+            return "Ask before executing any tool"
+        case .autoApproveRead:
+            return "Auto-approve file reads, grep, glob. Ask for writes and commands."
+        case .autoApproveAll:
+            return "Execute all tools without asking"
+        }
+    }
+
+    /// Set of tool names considered read-only
+    static let readOnlyTools: Set<String> = [
+        "file_operations",  // Only when operation=read
+        "grep",
+        "glob",
+        "code_navigation",
+        "workspace_analysis"
+    ]
+
+    /// Check if a tool call should be auto-approved under this policy
+    func shouldAutoApprove(toolName: String, parameters: [String: Any]) -> Bool {
+        switch self {
+        case .alwaysAsk:
+            return false
+        case .autoApproveAll:
+            return true
+        case .autoApproveRead:
+            // Special case: file_operations is only read-only for "read" operation
+            if toolName == "file_operations" {
+                let operation = parameters["operation"] as? String ?? ""
+                return operation == "read" || operation == "list"
+            }
+            return Self.readOnlyTools.contains(toolName)
+        }
+    }
+}

MLX Code/Models/ToolApproval.swift

@@ -0,0 +1,75 @@
+//
+//  ContextBudget.swift
+//  MLX Code
+//
+//  Token budget allocation for context-window-aware message assembly.
+//  Ensures system prompt, tools, recent messages, and project context
+//  all fit within the model's context window.
+//  Created on 2026-02-19.
+//
+
+import Foundation
+
+/// Manages token budget allocation across context components
+struct ContextBudget {
+    /// Total token budget (model's context window)
+    let totalBudget: Int
+
+    /// Fixed allocation for system prompt + tool descriptions
+    let systemPromptBudget: Int
+
+    /// Fixed allocation for few-shot examples
+    let fewShotBudget: Int
+
+    /// Reserved tokens for model's response output
+    let outputReservation: Int
+
+    /// Tokens available for conversation content
+    var conversationBudget: Int {
+        max(0, totalBudget - systemPromptBudget - fewShotBudget - outputReservation)
+    }
+
+    /// Budget for recent messages (70% of conversation budget)
+    var recentMessagesBudget: Int {
+        Int(Double(conversationBudget) * 0.7)
+    }
+
+    /// Budget for project context — file tree, recent files (20%)
+    var projectContextBudget: Int {
+        Int(Double(conversationBudget) * 0.2)
+    }
+
+    /// Budget for conversation summary of dropped messages (10%)
+    var summaryBudget: Int {
+        Int(Double(conversationBudget) * 0.1)
+    }
+
+    /// Creates a context budget for the given model
+    static func forModel(_ model: MLXModel?, daemonContextWindow: Int? = nil) -> ContextBudget {
+        // Prefer daemon-reported context window, then model config, then name-based heuristic
+        let contextWindow = daemonContextWindow
+            ?? model?.contextWindowSize
+            ?? detectContextWindow(modelName: model?.name ?? "")
+
+        return ContextBudget(
+            totalBudget: contextWindow,
+            systemPromptBudget: 500,
+            fewShotBudget: 150,
+            outputReservation: min(2048, contextWindow / 4)
+        )
+    }
+
+    /// Heuristic context window detection from model name
+    private static func detectContextWindow(modelName: String) -> Int {
+        let name = modelName.lowercased()
+        if name.contains("qwen") { return 32768 }
+        if name.contains("llama-3.1") || name.contains("llama 3.1") { return 131072 }
+        if name.contains("llama-3.2") || name.contains("llama 3.2") { return 8192 }
+        if name.contains("mistral") { return 32768 }
+        if name.contains("phi") { return 4096 }
+        if name.contains("deepseek") { return 16384 }
+        if name.contains("codellama") { return 16384 }
+        if name.contains("gemma") { return 8192 }
+        return 8192  // Conservative default
+    }
+}