Fix: Display original data:image URLs in rich display mode to maintain transparency

* initial terminal output chunking

* changeset

* Update src/core/task/index.ts

Co-authored-by: ellipsis-dev[bot] <65095814+ellipsis-dev[bot]@users.noreply.github.com>

* cleanup

---------

Co-authored-by: ellipsis-dev[bot] <65095814+ellipsis-dev[bot]@users.noreply.github.com>

.changeset/chatty-nails-sniff.md

@@ -1,5 +0,0 @@
----
-"claude-dev": minor
----
-
-better prompting around downloading mcp servers to not overwrite existing server configs

.changeset/cuddly-scissors-reflect.md

@@ -1,5 +0,0 @@
----
-"claude-dev": minor
----
-
-Added auto-approve options for edits/reads outside of the users workspace

.changeset/dirty-guests-shout.md

@@ -1,5 +0,0 @@
----
-"claude-dev": patch
----
-
-Improve diff editing animation and prompts for large files; show diff edits indicator next to file path

.changeset/eleven-comics-roll.md

@@ -1,5 +0,0 @@
----
-"claude-dev": minor
----
-
-Add button to open config to MCP toggle modal

.changeset/enhance-ollama-provider.md

@@ -1,5 +0,0 @@
----
-"claude-dev": minor
----
-
-Enhance Ollama provider with retry mechanism, timeout handling, and improved error handling. This change adds robust error handling, automatic retries, timeout handling, and improved stream processing to the Ollama provider, making it more reliable and preventing the infinite "thinking" problem. Tests are now skipped if Ollama is not running locally.

.changeset/fast-readers-care.md

@@ -0,0 +1,5 @@
+---
+"claude-dev": minor
+---
+
+Added chunking to terminal outputs

.changeset/gorgeous-bees-talk.md

@@ -1,5 +0,0 @@
----
-"claude-dev": patch
----
-
-Added debounce for checkmark expanded ui

.changeset/great-onions-jump.md

@@ -1,5 +0,0 @@
----
-"claude-dev": minor
----
-
-Add more models support for doubao

.changeset/green-cows-occur.md

@@ -1,5 +0,0 @@
----
-"claude-dev": minor
----
-
-Switch clinerules file to a directory

.changeset/kind-terms-dance.md

@@ -1,5 +0,0 @@
----
-"claude-dev": patch
----
-
-Refactor: Remove cline rules functions from task/index.ts

.changeset/late-mice-double.md

@@ -0,0 +1,5 @@
+---
+"claude-dev": patch
+---
+
+gRPC over vscode message bus to make messaging better

.changeset/nasty-gifts-look.md

@@ -0,0 +1,5 @@
+---
+"claude-dev": patch
+---
+
+Allow user to modify a previous message to continue chat

.changeset/nasty-months-wave.md

@@ -0,0 +1,5 @@
+---
+"claude-dev": patch
+---
+
+Add command to focus chat input

.changeset/neat-wolves-grab.md

@@ -1,5 +0,0 @@
----
-"claude-dev": patch
----
-
-Added telemetry for model favoriting

.changeset/nine-walls-report.md

@@ -1,5 +0,0 @@
----
-"claude-dev": minor
----
-
-Add reasoning effort to grok 3 mini

.changeset/pretty-eels-breathe.md

@@ -0,0 +1,5 @@
+---
+"claude-dev": patch
+---
+
+fix vertexai token count

.changeset/rare-starfishes-design.md

@@ -0,0 +1,5 @@
+---
+"claude-dev": patch
+---
+
+Allow user to send context with an option selection

.changeset/rich-pens-buy.md

@@ -0,0 +1,5 @@
+---
+"claude-dev": patch
+---
+
+Safer check for xAI reasoning content

.changeset/rude-apes-behave.md

@@ -1,5 +0,0 @@
----
-"claude-dev": patch
----
-
-Add model info to the Requesty provider

.changeset/rude-rice-raise.md

@@ -1,5 +0,0 @@
----
-"claude-dev": patch
----
-
-refactor & perf of file ServerRow.tsx

.changeset/short-donuts-search.md

@@ -1,5 +0,0 @@
----
-"claude-dev": patch
----
-
-remove unused import of file `BrowserSettingsMenu.tsx`

.changeset/six-birds-move.md

@@ -0,0 +1,5 @@
+---
+"claude-dev": minor
+---
+
+Add an edit button to the Cline Rules toggle modal to open rule files

.changeset/sixty-jokes-hope.md

@@ -1,5 +0,0 @@
----
-"claude-dev": minor
----
-
-Stream Grok 3 mini reasoning tokens

.changeset/slimy-jeans-dance.md

@@ -0,0 +1,5 @@
+---
+"claude-dev": patch
+---
+
+MCP ImageContent as tool result

.changeset/tasty-plums-laugh.md

@@ -1,5 +0,0 @@
----
-"claude-dev": patch
----
-
-refactor & perf of file `ThinkingBudgetSlider.tsx`

.changeset/ten-bananas-begin.md

@@ -2,4 +2,4 @@
 "claude-dev": patch
 ---
 
-Remove WeakRef usage
+Readme update

.changeset/tough-clouds-design.md

@@ -1,5 +0,0 @@
----
-"claude-dev": patch
----
-
-refactor: type improvement of file `ChatTextArea.tsx`

.changeset/twelve-eyes-teach.md

@@ -0,0 +1,5 @@
+---
+"claude-dev": patch
+---
+
+Add OpenAI o3 & 4o-mini

.changeset/twelve-meals-applaud.md

@@ -0,0 +1,5 @@
+---
+"claude-dev": minor
+---
+
+Add modal UI for toggling Cline Rules

.changeset/twelve-rocks-drum.md

@@ -1,5 +0,0 @@
----
-"claude-dev": minor
----
-
-Adds favorite toggles for models when using the Cline & OpenRouter providers

.codespellrc

@@ -0,0 +1,6 @@
+[codespell]
+# Ref: https://github.com/codespell-project/codespell#using-a-config-file
+skip = .git*,*.svg,package-lock.json,*.css,.codespellrc,locales
+check-hidden = true
+ignore-regex = (\b(optIn|isTaller)\b|https://\S+)
+# ignore-words-list =

.github/workflows/codespell.yml

@@ -0,0 +1,25 @@
+# Codespell configuration is within .codespellrc
+---
+name: Codespell
+
+on:
+    push:
+        branches: [main]
+    pull_request:
+        branches: [main]
+
+permissions:
+    contents: read
+
+jobs:
+    codespell:
+        name: Check for spelling errors
+        runs-on: ubuntu-latest
+
+        steps:
+            - name: Checkout
+              uses: actions/checkout@v4
+            - name: Annotate locations with typos
+              uses: codespell-project/codespell-problem-matcher@v1
+            - name: Codespell
+              uses: codespell-project/actions-codespell@v2

CHANGELOG.md

@@ -1,5 +1,34 @@
 # Changelog
 
+## [3.12.3]
+
+-   Add copy button to MermaidBlock component (Thanks @cacosub7!)
+-   Add the ability to fetch from global cline rules files
+-   Add icon to indicate when a file outside of the users workspace is edited
+
+## [3.12.2]
+
+-   Add gpt-4.1
+
+## [3.12.1]
+
+-   Use visual checkpoint indicator to make it clear when checkpoints are created
+-   Big shoutout to @samuel871211 for numerous code quality improvements, refactoring contributions, and webview performance improvements!
+-   Use improved context manager
+
+## [3.12.0]
+
+-   Add favorite toggles for models when using the Cline & OpenRouter providers
+-   Add auto-approve options for edits/reads outside of the workspace
+-   Improve diff editing animation for large files
+-   Add indicator showing number of diff edits when Cline edits a file
+-   Add streaming support and reasoning effort option to xAI's Grok 3 Mini
+-   Add settings button to MCP popover to easily modify installed servers
+-   Fix bug where browser tool actions would show unparsed results in the chat view
+-   Fix issue with new checkpoints popover hiding too quickly
+-   Fix duplicate checkpoints bug
+-   Improve Ollama provider with retry mechanism, timeout handling, and improved error handling (thanks suvarchal!)
+
 ## [3.11.0]
 
 -   Redesign checkpoint UI to declutter chat view by using a subtle indicator line that expands to a popover on hover, with a new date indicator for when it was created
@@ -29,7 +58,7 @@
 -   Add recommended models for Cline provider
 -   Add ability to detect when user edits files manually so Cline knows to re-read, leading to reduced diff edit errors
 -   Add improvements to file mention searching for faster searching
--   Add scoring logic to file mentions to sort and exlcude results based on relevance
+-   Add scoring logic to file mentions to sort and exclude results based on relevance
 -   Add Support for Bytedance Doubao (Thanks Tunixer!)
 -   Fix to prevent duplicate BOM (Thanks bamps53!)
 
@@ -507,7 +536,7 @@
 -   Adds "Always allow read-only operations" setting to let Claude read files and view directories without needing approval (off by default)
 -   Implement sliding window context management to keep tasks going past 200k tokens
 -   Adds Google Cloud Vertex AI support and updates Claude 3.5 Sonnet max output to 8192 tokens for all providers.
--   Improves system prompt to gaurd against lazy edits (less "//rest of code here")
+-   Improves system prompt to guard against lazy edits (less "//rest of code here")
 
 ## [1.3.0]
 

evals/.gitignore

@@ -0,0 +1,3 @@
+repositories
+
+results/evals.db

evals/README.md

@@ -0,0 +1,162 @@
+# Cline Evaluation System
+
+This directory contains the evaluation system for benchmarking Cline against various coding evaluation frameworks.
+
+## Overview
+
+The Cline Evaluation System allows you to:
+
+1. Run Cline against standardized coding benchmarks
+2. Collect comprehensive metrics on performance
+3. Generate detailed reports on evaluation results
+4. Compare performance across different models and benchmarks
+
+## Architecture
+
+The evaluation system consists of two main components:
+
+1. **Test Server**: Enhanced HTTP server in `src/services/test/TestServer.ts` that provides detailed task results
+2. **CLI Tool**: Command-line interface in `evals/cli/` for orchestrating evaluations
+
+## Directory Structure
+
+```
+cline-repo/
+├── src/
+│   ├── services/
+│   │   ├── test/
+│   │   │   ├── TestServer.ts         # Enhanced HTTP server for task execution
+│   │   │   ├── GitHelper.ts          # Git utilities for file tracking
+│   │   │   └── ...
+│   │   └── ...
+│   └── ...
+├── evals/                            # Main directory for evaluation system
+│   ├── cli/                          # CLI tool for orchestrating evaluations
+│   │   ├── src/
+│   │   │   ├── index.ts              # CLI entry point
+│   │   │   ├── commands/             # CLI commands (setup, run, report)
+│   │   │   ├── adapters/             # Benchmark adapters
+│   │   │   ├── db/                   # Database management
+│   │   │   └── utils/                # Utility functions
+│   │   ├── package.json
+│   │   └── tsconfig.json
+│   ├── repositories/                 # Cloned benchmark repositories
+│   │   ├── exercism/                 # Modified Exercism (from pashpashpash/evals)
+│   │   ├── swe-bench/                # SWE-Bench repository
+│   │   ├── swelancer/                # SWELancer repository
+│   │   └── multi-swe/                # Multi-SWE-Bench repository
+│   ├── results/                      # Evaluation results storage
+│   │   ├── runs/                     # Individual run results
+│   │   └── reports/                  # Generated reports
+│   └── README.md                     # This file
+└── ...
+```
+
+## Getting Started
+
+### Prerequisites
+
+- Node.js 16+
+- VSCode with Cline extension installed
+- Git
+
+### Installation
+
+1. Build the CLI tool:
+
+```bash
+cd evals/cli
+npm install
+npm run build
+```
+
+### Usage
+
+#### Setting Up Benchmarks
+
+```bash
+cd evals/cli
+node dist/index.js setup
+```
+
+This will clone and set up all benchmark repositories. You can specify specific benchmarks:
+
+```bash
+node dist/index.js setup --benchmarks exercism
+```
+
+#### Running Evaluations
+
+```bash
+node dist/index.js run --model claude-3-opus-20240229 --benchmark exercism
+```
+
+Options:
+- `--model`: The model to evaluate (default: claude-3-opus-20240229)
+- `--benchmark`: Specific benchmark to run (default: all)
+- `--count`: Number of tasks to run (default: all)
+
+#### Generating Reports
+
+```bash
+node dist/index.js report
+```
+
+Options:
+- `--format`: Report format (json, markdown) (default: markdown)
+- `--output`: Output path for the report
+
+## Benchmarks
+
+### Exercism
+
+Modified Exercism exercises from the [pashpashpash/evals](https://github.com/pashpashpash/evals) repository. These are small, focused programming exercises in various languages.
+
+### SWE-Bench (Coming Soon)
+
+Real-world software engineering tasks from the [SWE-bench](https://github.com/SWE-bench/SWE-bench) repository.
+
+### SWELancer (Coming Soon)
+
+Freelance-style programming tasks from the SWELancer benchmark.
+
+### Multi-SWE-Bench (Coming Soon)
+
+Multi-file software engineering tasks from the Multi-SWE-Bench repository.
+
+## Metrics
+
+The evaluation system collects the following metrics:
+
+- **Token Usage**: Input and output tokens
+- **Cost**: Estimated cost of API calls
+- **Duration**: Time taken to complete tasks
+- **Tool Usage**: Number of tool calls and failures
+- **Success Rate**: Percentage of tasks completed successfully
+- **Functional Correctness**: Percentage of tests passed
+
+## Reports
+
+Reports are generated in Markdown or JSON format and include:
+
+- Overall summary
+- Benchmark-specific results
+- Model-specific results
+- Tool usage statistics
+- Charts and visualizations
+
+## Development
+
+### Adding a New Benchmark
+
+1. Create a new adapter in `evals/cli/src/adapters/`
+2. Implement the `BenchmarkAdapter` interface
+3. Register the adapter in `evals/cli/src/adapters/index.ts`
+
+### Extending Metrics
+
+To add new metrics:
+
+1. Update the database schema in `evals/cli/src/db/schema.ts`
+2. Add collection logic in `evals/cli/src/utils/results.ts`
+3. Update report generation in `evals/cli/src/commands/report.ts`

evals/cli/package.json

@@ -0,0 +1,39 @@
+{
+	"name": "cline-evaluation-cli",
+	"version": "0.1.0",
+	"description": "CLI tool for orchestrating Cline evaluations across multiple benchmarks",
+	"main": "dist/index.js",
+	"scripts": {
+		"build": "tsc",
+		"start": "node dist/index.js",
+		"dev": "ts-node src/index.ts",
+		"test": "echo \"Error: no test specified\" && exit 1"
+	},
+	"keywords": [
+		"cline",
+		"evaluation",
+		"benchmark"
+	],
+	"author": "",
+	"license": "MIT",
+	"dependencies": {
+		"better-sqlite3": "^8.0.0",
+		"chalk": "^4.1.2",
+		"commander": "^9.4.1",
+		"execa": "^5.1.1",
+		"node-fetch": "^2.7.0",
+		"ora": "^5.4.1",
+		"sqlite": "^4.1.2",
+		"uuid": "^9.0.0",
+		"yargs": "^17.6.2"
+	},
+	"devDependencies": {
+		"@types/better-sqlite3": "^7.6.3",
+		"@types/node": "^18.11.18",
+		"@types/node-fetch": "^2.6.12",
+		"@types/uuid": "^9.0.0",
+		"@types/yargs": "^17.0.19",
+		"ts-node": "^10.9.1",
+		"typescript": "^4.9.4"
+	}
+}

evals/cli/src/adapters/exercism.ts

@@ -0,0 +1,190 @@
+import * as path from "path"
+import * as fs from "fs"
+import execa from "execa"
+import { BenchmarkAdapter, Task, VerificationResult } from "./types"
+
+const EVALS_DIR = path.resolve(__dirname, "../../../")
+
+/**
+ * Adapter for the modified Exercism benchmark
+ */
+export class ExercismAdapter implements BenchmarkAdapter {
+	name = "exercism"
+
+	/**
+	 * Set up the Exercism benchmark repository
+	 */
+	async setup(): Promise<void> {
+		// Clone repository if needed
+		const exercismDir = path.join(EVALS_DIR, "repositories", "exercism")
+
+		if (!fs.existsSync(exercismDir)) {
+			console.log(`Cloning Exercism repository to ${exercismDir}...`)
+			await execa("git", ["clone", "https://github.com/pashpashpash/evals.git", exercismDir])
+			console.log("Exercism repository cloned successfully")
+		} else {
+			console.log(`Exercism repository already exists at ${exercismDir}`)
+
+			// Pull latest changes
+			console.log("Pulling latest changes...")
+			await execa("git", ["pull"], { cwd: exercismDir })
+			console.log("Repository updated successfully")
+		}
+	}
+
+	/**
+	 * List all available tasks in the Exercism benchmark
+	 */
+	async listTasks(): Promise<Task[]> {
+		const tasks: Task[] = []
+		const exercisesDir = path.join(EVALS_DIR, "repositories", "exercism")
+
+		// Ensure the repository exists
+		if (!fs.existsSync(exercisesDir)) {
+			throw new Error(`Exercism repository not found at ${exercisesDir}. Run setup first.`)
+		}
+
+		// Read language directories
+		const languages = fs
+			.readdirSync(exercisesDir)
+			.filter((dir) => fs.statSync(path.join(exercisesDir, dir)).isDirectory())
+			.filter((dir) => !dir.startsWith(".") && !["node_modules", ".git"].includes(dir))
+
+		for (const language of languages) {
+			const languageDir = path.join(exercisesDir, language)
+
+			// Read exercise directories
+			const exercises = fs.readdirSync(languageDir).filter((dir) => fs.statSync(path.join(languageDir, dir)).isDirectory())
+
+			for (const exercise of exercises) {
+				const exerciseDir = path.join(languageDir, exercise)
+
+				// Read instructions
+				let description = ""
+				const instructionsPath = path.join(exerciseDir, "docs", "instructions.md")
+				if (fs.existsSync(instructionsPath)) {
+					description = fs.readFileSync(instructionsPath, "utf-8")
+				}
+
+				// Determine test commands based on language
+				let testCommands: string[] = []
+				switch (language) {
+					case "javascript":
+						testCommands = ["npm install", "npm test"]
+						break
+					case "python":
+						testCommands = ["python -m pytest -o markers=task *_test.py"]
+						break
+					case "go":
+						testCommands = ["go test"]
+						break
+					case "java":
+						testCommands = ["./gradlew test"]
+						break
+					case "rust":
+						testCommands = ["cargo test"]
+						break
+					default:
+						testCommands = []
+				}
+
+				tasks.push({
+					id: `exercism-${language}-${exercise}`,
+					name: exercise,
+					description,
+					workspacePath: exerciseDir,
+					setupCommands: [],
+					verificationCommands: testCommands,
+					metadata: {
+						language,
+						type: "exercism",
+					},
+				})
+			}
+		}
+
+		return tasks
+	}
+
+	/**
+	 * Prepare a specific task for execution
+	 * @param taskId The ID of the task to prepare
+	 */
+	async prepareTask(taskId: string): Promise<Task> {
+		const tasks = await this.listTasks()
+		const task = tasks.find((t) => t.id === taskId)
+
+		if (!task) {
+			throw new Error(`Task ${taskId} not found`)
+		}
+
+		// Check if Git repository is already initialized
+		const gitDirExists = fs.existsSync(path.join(task.workspacePath, ".git"))
+
+		try {
+			// Initialize Git repository if needed
+			if (!gitDirExists) {
+				await execa("git", ["init"], { cwd: task.workspacePath })
+			}
+
+			// Create a dummy file to ensure there's something to commit
+			const dummyFilePath = path.join(task.workspacePath, ".eval-timestamp")
+			fs.writeFileSync(dummyFilePath, new Date().toISOString())
+
+			// Add all files and commit
+			await execa("git", ["add", "."], { cwd: task.workspacePath })
+
+			try {
+				await execa("git", ["commit", "-m", "Initial commit"], { cwd: task.workspacePath })
+			} catch (error: any) {
+				// If commit fails because there are no changes, that's okay
+				if (!error.stderr?.includes("nothing to commit")) {
+					throw error
+				}
+			}
+		} catch (error: any) {
+			console.warn(`Warning: Git operations failed: ${error.message}`)
+			console.warn("Continuing without Git initialization")
+		}
+
+		return task
+	}
+
+	/**
+	 * Verify the result of a task execution
+	 * @param task The task that was executed
+	 * @param result The result of the task execution
+	 */
+	async verifyResult(task: Task, result: any): Promise<VerificationResult> {
+		// Run verification commands
+		let success = true
+		let output = ""
+
+		for (const command of task.verificationCommands) {
+			try {
+				const [cmd, ...args] = command.split(" ")
+				const { stdout } = await execa(cmd, args, { cwd: task.workspacePath })
+				output += stdout + "\n"
+			} catch (error: any) {
+				success = false
+				if (error.stdout) output += error.stdout + "\n"
+				if (error.stderr) output += error.stderr + "\n"
+			}
+		}
+
+		// Parse test results
+		const testsPassed = (output.match(/PASS/g) || []).length
+		const testsFailed = (output.match(/FAIL/g) || []).length
+		const testsTotal = testsPassed + testsFailed
+
+		return {
+			success,
+			metrics: {
+				testsPassed,
+				testsFailed,
+				testsTotal,
+				functionalCorrectness: testsTotal > 0 ? testsPassed / testsTotal : 0,
+			},
+		}
+	}
+}

evals/cli/src/adapters/index.ts

@@ -0,0 +1,47 @@
+import { BenchmarkAdapter } from "./types"
+import { ExercismAdapter } from "./exercism"
+import { SWEBenchAdapter } from "./swe-bench"
+import { SWELancerAdapter } from "./swelancer"
+import { MultiSWEAdapter } from "./multi-swe"
+
+// Registry of all available adapters
+const adapters: Record<string, BenchmarkAdapter> = {
+	// Exercism is the primary adapter with real implementation
+	exercism: new ExercismAdapter(),
+
+	// Dummy adapters for testing
+	"swe-bench": new SWEBenchAdapter(),
+	swelancer: new SWELancerAdapter(),
+	"multi-swe": new MultiSWEAdapter(),
+}
+
+/**
+ * Get a specific adapter by name
+ * @param name The name of the adapter to get
+ * @returns The requested adapter
+ * @throws Error if the adapter is not found
+ */
+export function getAdapter(name: string): BenchmarkAdapter {
+	const adapter = adapters[name]
+	if (!adapter) {
+		throw new Error(`Adapter for benchmark '${name}' not found`)
+	}
+	return adapter
+}
+
+/**
+ * Get all available adapters
+ * @returns Array of all registered adapters
+ */
+export function getAllAdapters(): BenchmarkAdapter[] {
+	return Object.values(adapters)
+}
+
+/**
+ * Register a new adapter
+ * @param name The name to register the adapter under
+ * @param adapter The adapter to register
+ */
+export function registerAdapter(name: string, adapter: BenchmarkAdapter): void {
+	adapters[name] = adapter
+}

evals/cli/src/adapters/multi-swe.ts

@@ -0,0 +1,192 @@
+import * as path from "path"
+import * as fs from "fs"
+import execa from "execa"
+import { BenchmarkAdapter, Task, VerificationResult } from "./types"
+
+const EVALS_DIR = path.resolve(__dirname, "../../../")
+
+/**
+ * Dummy adapter for the Multi-SWE-Bench benchmark
+ */
+export class MultiSWEAdapter implements BenchmarkAdapter {
+	name = "multi-swe"
+
+	/**
+	 * Set up the Multi-SWE-Bench benchmark repository (dummy implementation)
+	 */
+	async setup(): Promise<void> {
+		console.log("Multi-SWE-Bench dummy setup completed")
+
+		// Create repositories directory if it doesn't exist
+		const repoDir = path.join(EVALS_DIR, "repositories", "multi-swe")
+		if (!fs.existsSync(repoDir)) {
+			fs.mkdirSync(repoDir, { recursive: true })
+			console.log(`Created dummy Multi-SWE-Bench directory at ${repoDir}`)
+		}
+	}
+
+	/**
+	 * List all available tasks in the Multi-SWE-Bench benchmark (dummy implementation)
+	 */
+	async listTasks(): Promise<Task[]> {
+		return [
+			{
+				id: "multi-swe-task-1",
+				name: "Multi-Language API Integration",
+				description:
+					"Implement a system that integrates a Python backend with a TypeScript frontend and a Rust processing service.",
+				workspacePath: path.join(EVALS_DIR, "repositories", "multi-swe"),
+				setupCommands: [],
+				verificationCommands: [],
+				metadata: {
+					languages: ["python", "typescript", "rust"],
+					complexity: "high",
+					type: "multi-swe",
+				},
+			},
+			{
+				id: "multi-swe-task-2",
+				name: "Cross-Platform Mobile App",
+				description: "Create a cross-platform mobile app using React Native with native modules in Swift and Kotlin.",
+				workspacePath: path.join(EVALS_DIR, "repositories", "multi-swe"),
+				setupCommands: [],
+				verificationCommands: [],
+				metadata: {
+					languages: ["javascript", "swift", "kotlin"],
+					complexity: "medium",
+					type: "multi-swe",
+				},
+			},
+			{
+				id: "multi-swe-task-3",
+				name: "Microservice Architecture",
+				description: "Design and implement a microservice architecture with services written in Go, Node.js, and Java.",
+				workspacePath: path.join(EVALS_DIR, "repositories", "multi-swe"),
+				setupCommands: [],
+				verificationCommands: [],
+				metadata: {
+					languages: ["go", "javascript", "java"],
+					complexity: "high",
+					type: "multi-swe",
+				},
+			},
+		]
+	}
+
+	/**
+	 * Prepare a specific task for execution (dummy implementation)
+	 * @param taskId The ID of the task to prepare
+	 */
+	async prepareTask(taskId: string): Promise<Task> {
+		const tasks = await this.listTasks()
+		const task = tasks.find((t) => t.id === taskId)
+
+		if (!task) {
+			throw new Error(`Task ${taskId} not found`)
+		}
+
+		// Create a dummy workspace for the task
+		const taskDir = path.join(task.workspacePath, taskId)
+		if (!fs.existsSync(taskDir)) {
+			fs.mkdirSync(taskDir, { recursive: true })
+
+			// Create a dummy file for the task
+			fs.writeFileSync(
+				path.join(taskDir, "README.md"),
+				`# ${task.name}\n\n${task.description}\n\nThis is a dummy task for testing purposes.`,
+			)
+
+			// Create additional dummy files based on task type
+			if (task.id === "multi-swe-task-1") {
+				// Python backend
+				fs.mkdirSync(path.join(taskDir, "backend"), { recursive: true })
+				fs.writeFileSync(
+					path.join(taskDir, "backend", "app.py"),
+					`# TODO: Implement Python backend\nfrom flask import Flask\n\napp = Flask(__name__)\n\n@app.route('/')\ndef hello():\n    return "Hello, World!"\n`,
+				)
+
+				// TypeScript frontend
+				fs.mkdirSync(path.join(taskDir, "frontend"), { recursive: true })
+				fs.writeFileSync(
+					path.join(taskDir, "frontend", "app.ts"),
+					`// TODO: Implement TypeScript frontend\nconsole.log('Frontend starting...');\n`,
+				)
+
+				// Rust processing service
+				fs.mkdirSync(path.join(taskDir, "processor"), { recursive: true })
+				fs.writeFileSync(
+					path.join(taskDir, "processor", "main.rs"),
+					`// TODO: Implement Rust processing service\nfn main() {\n    println!("Processor starting...");\n}\n`,
+				)
+			} else if (task.id === "multi-swe-task-2") {
+				// React Native app
+				fs.mkdirSync(path.join(taskDir, "app"), { recursive: true })
+				fs.writeFileSync(
+					path.join(taskDir, "app", "App.js"),
+					`// TODO: Implement React Native app\nimport React from 'react';\nimport { View, Text } from 'react-native';\n\nexport default function App() {\n  return (\n    <View>\n      <Text>Hello, World!</Text>\n    </View>\n  );\n}\n`,
+				)
+
+				// Swift native module
+				fs.mkdirSync(path.join(taskDir, "ios"), { recursive: true })
+				fs.writeFileSync(
+					path.join(taskDir, "ios", "NativeModule.swift"),
+					`// TODO: Implement Swift native module\nimport Foundation\n\n@objc(NativeModule)\nclass NativeModule: NSObject {\n  @objc\n  func hello() -> String {\n    return "Hello from Swift"\n  }\n}\n`,
+				)
+
+				// Kotlin native module
+				fs.mkdirSync(path.join(taskDir, "android"), { recursive: true })
+				fs.writeFileSync(
+					path.join(taskDir, "android", "NativeModule.kt"),
+					`// TODO: Implement Kotlin native module\npackage com.example.app\n\nclass NativeModule {\n  fun hello(): String {\n    return "Hello from Kotlin"\n  }\n}\n`,
+				)
+			} else if (task.id === "multi-swe-task-3") {
+				// Go service
+				fs.mkdirSync(path.join(taskDir, "service-go"), { recursive: true })
+				fs.writeFileSync(
+					path.join(taskDir, "service-go", "main.go"),
+					`// TODO: Implement Go service\npackage main\n\nimport "fmt"\n\nfunc main() {\n\tfmt.Println("Go service starting...")\n}\n`,
+				)
+
+				// Node.js service
+				fs.mkdirSync(path.join(taskDir, "service-node"), { recursive: true })
+				fs.writeFileSync(
+					path.join(taskDir, "service-node", "server.js"),
+					`// TODO: Implement Node.js service\nconsole.log('Node.js service starting...');\n`,
+				)
+
+				// Java service
+				fs.mkdirSync(path.join(taskDir, "service-java"), { recursive: true })
+				fs.writeFileSync(
+					path.join(taskDir, "service-java", "Main.java"),
+					`// TODO: Implement Java service\npublic class Main {\n    public static void main(String[] args) {\n        System.out.println("Java service starting...");\n    }\n}\n`,
+				)
+			}
+		}
+
+		// Update the task's workspace path to the task-specific directory
+		return {
+			...task,
+			workspacePath: taskDir,
+		}
+	}
+
+	/**
+	 * Verify the result of a task execution (dummy implementation)
+	 * @param task The task that was executed
+	 * @param result The result of the task execution
+	 */
+	async verifyResult(task: Task, result: any): Promise<VerificationResult> {
+		// Always return success for dummy implementation
+		return {
+			success: true,
+			metrics: {
+				testsPassed: 1,
+				testsFailed: 0,
+				testsTotal: 1,
+				functionalCorrectness: 1.0,
+				crossLanguageIntegration: 0.9, // Dummy metric specific to Multi-SWE
+				architectureQuality: 0.85, // Dummy metric specific to Multi-SWE
+			},
+		}
+	}
+}

evals/cli/src/adapters/swe-bench.ts

@@ -0,0 +1,125 @@
+import * as path from "path"
+import * as fs from "fs"
+import execa from "execa"
+import { BenchmarkAdapter, Task, VerificationResult } from "./types"
+
+const EVALS_DIR = path.resolve(__dirname, "../../../")
+
+/**
+ * Dummy adapter for the SWE-Bench benchmark
+ */
+export class SWEBenchAdapter implements BenchmarkAdapter {
+	name = "swe-bench"
+
+	/**
+	 * Set up the SWE-Bench benchmark repository (dummy implementation)
+	 */
+	async setup(): Promise<void> {
+		console.log("SWE-Bench dummy setup completed")
+
+		// Create repositories directory if it doesn't exist
+		const repoDir = path.join(EVALS_DIR, "repositories", "swe-bench")
+		if (!fs.existsSync(repoDir)) {
+			fs.mkdirSync(repoDir, { recursive: true })
+			console.log(`Created dummy SWE-Bench directory at ${repoDir}`)
+		}
+	}
+
+	/**
+	 * List all available tasks in the SWE-Bench benchmark (dummy implementation)
+	 */
+	async listTasks(): Promise<Task[]> {
+		return [
+			{
+				id: "swe-bench-task-1",
+				name: "Fix React Component Bug",
+				description: "Fix a bug in a React component where the state is not properly updated.",
+				workspacePath: path.join(EVALS_DIR, "repositories", "swe-bench"),
+				setupCommands: [],
+				verificationCommands: [],
+				metadata: {
+					repository: "facebook/react",
+					issue: "#12345",
+					type: "swe-bench",
+				},
+			},
+			{
+				id: "swe-bench-task-2",
+				name: "Optimize Database Query",
+				description: "Optimize a slow database query in a Django application.",
+				workspacePath: path.join(EVALS_DIR, "repositories", "swe-bench"),
+				setupCommands: [],
+				verificationCommands: [],
+				metadata: {
+					repository: "django/django",
+					issue: "#6789",
+					type: "swe-bench",
+				},
+			},
+			{
+				id: "swe-bench-task-3",
+				name: "Fix Memory Leak",
+				description: "Fix a memory leak in a Node.js application.",
+				workspacePath: path.join(EVALS_DIR, "repositories", "swe-bench"),
+				setupCommands: [],
+				verificationCommands: [],
+				metadata: {
+					repository: "nodejs/node",
+					issue: "#9876",
+					type: "swe-bench",
+				},
+			},
+		]
+	}
+
+	/**
+	 * Prepare a specific task for execution (dummy implementation)
+	 * @param taskId The ID of the task to prepare
+	 */
+	async prepareTask(taskId: string): Promise<Task> {
+		const tasks = await this.listTasks()
+		const task = tasks.find((t) => t.id === taskId)
+
+		if (!task) {
+			throw new Error(`Task ${taskId} not found`)
+		}
+
+		// Create a dummy workspace for the task
+		const taskDir = path.join(task.workspacePath, taskId)
+		if (!fs.existsSync(taskDir)) {
+			fs.mkdirSync(taskDir, { recursive: true })
+
+			// Create a dummy file for the task
+			fs.writeFileSync(
+				path.join(taskDir, "README.md"),
+				`# ${task.name}\n\n${task.description}\n\nThis is a dummy task for testing purposes.`,
+			)
+		}
+
+		// Update the task's workspace path to the task-specific directory
+		return {
+			...task,
+			workspacePath: taskDir,
+		}
+	}
+
+	/**
+	 * Verify the result of a task execution (dummy implementation)
+	 * @param task The task that was executed
+	 * @param result The result of the task execution
+	 */
+	async verifyResult(task: Task, result: any): Promise<VerificationResult> {
+		// Always return success for dummy implementation
+		return {
+			success: true,
+			metrics: {
+				testsPassed: 1,
+				testsFailed: 0,
+				testsTotal: 1,
+				functionalCorrectness: 1.0,
+				performanceImprovement: 0.25, // Dummy metric specific to SWE-Bench
+				codeQuality: 0.9, // Dummy metric specific to SWE-Bench
+			},
+		}
+	}
+}

evals/cli/src/adapters/swelancer.ts

@@ -0,0 +1,143 @@
+import * as path from "path"
+import * as fs from "fs"
+import execa from "execa"
+import { BenchmarkAdapter, Task, VerificationResult } from "./types"
+
+const EVALS_DIR = path.resolve(__dirname, "../../../")
+
+/**
+ * Dummy adapter for the SWELancer benchmark
+ */
+export class SWELancerAdapter implements BenchmarkAdapter {
+	name = "swelancer"
+
+	/**
+	 * Set up the SWELancer benchmark repository (dummy implementation)
+	 */
+	async setup(): Promise<void> {
+		console.log("SWELancer dummy setup completed")
+
+		// Create repositories directory if it doesn't exist
+		const repoDir = path.join(EVALS_DIR, "repositories", "swelancer")
+		if (!fs.existsSync(repoDir)) {
+			fs.mkdirSync(repoDir, { recursive: true })
+			console.log(`Created dummy SWELancer directory at ${repoDir}`)
+		}
+	}
+
+	/**
+	 * List all available tasks in the SWELancer benchmark (dummy implementation)
+	 */
+	async listTasks(): Promise<Task[]> {
+		return [
+			{
+				id: "swelancer-task-1",
+				name: "Create Landing Page",
+				description: "Create a responsive landing page for a new product using HTML, CSS, and JavaScript.",
+				workspacePath: path.join(EVALS_DIR, "repositories", "swelancer"),
+				setupCommands: [],
+				verificationCommands: [],
+				metadata: {
+					client: "TechStartup Inc.",
+					difficulty: "medium",
+					type: "swelancer",
+				},
+			},
+			{
+				id: "swelancer-task-2",
+				name: "Build REST API",
+				description: "Create a RESTful API for a blog application using Node.js and Express.",
+				workspacePath: path.join(EVALS_DIR, "repositories", "swelancer"),
+				setupCommands: [],
+				verificationCommands: [],
+				metadata: {
+					client: "BlogCo",
+					difficulty: "hard",
+					type: "swelancer",
+				},
+			},
+			{
+				id: "swelancer-task-3",
+				name: "Fix CSS Layout Issues",
+				description: "Fix layout issues in a responsive website across different screen sizes.",
+				workspacePath: path.join(EVALS_DIR, "repositories", "swelancer"),
+				setupCommands: [],
+				verificationCommands: [],
+				metadata: {
+					client: "DesignAgency",
+					difficulty: "easy",
+					type: "swelancer",
+				},
+			},
+		]
+	}
+
+	/**
+	 * Prepare a specific task for execution (dummy implementation)
+	 * @param taskId The ID of the task to prepare
+	 */
+	async prepareTask(taskId: string): Promise<Task> {
+		const tasks = await this.listTasks()
+		const task = tasks.find((t) => t.id === taskId)
+
+		if (!task) {
+			throw new Error(`Task ${taskId} not found`)
+		}
+
+		// Create a dummy workspace for the task
+		const taskDir = path.join(task.workspacePath, taskId)
+		if (!fs.existsSync(taskDir)) {
+			fs.mkdirSync(taskDir, { recursive: true })
+
+			// Create a dummy file for the task
+			fs.writeFileSync(
+				path.join(taskDir, "README.md"),
+				`# ${task.name}\n\n${task.description}\n\nThis is a dummy task for testing purposes.`,
+			)
+
+			// Create additional dummy files based on task type
+			if (task.id === "swelancer-task-1") {
+				fs.writeFileSync(
+					path.join(taskDir, "index.html"),
+					`<!DOCTYPE html>\n<html>\n<head>\n  <title>Landing Page</title>\n</head>\n<body>\n  <!-- TODO: Implement landing page -->\n</body>\n</html>`,
+				)
+			} else if (task.id === "swelancer-task-2") {
+				fs.writeFileSync(
+					path.join(taskDir, "server.js"),
+					`// TODO: Implement REST API\nconsole.log('Server starting...');`,
+				)
+			} else if (task.id === "swelancer-task-3") {
+				fs.writeFileSync(
+					path.join(taskDir, "styles.css"),
+					`/* TODO: Fix layout issues */\nbody {\n  margin: 0;\n  padding: 0;\n}`,
+				)
+			}
+		}
+
+		// Update the task's workspace path to the task-specific directory
+		return {
+			...task,
+			workspacePath: taskDir,
+		}
+	}
+
+	/**
+	 * Verify the result of a task execution (dummy implementation)
+	 * @param task The task that was executed
+	 * @param result The result of the task execution
+	 */
+	async verifyResult(task: Task, result: any): Promise<VerificationResult> {
+		// Always return success for dummy implementation
+		return {
+			success: true,
+			metrics: {
+				testsPassed: 1,
+				testsFailed: 0,
+				testsTotal: 1,
+				functionalCorrectness: 1.0,
+				clientSatisfaction: 0.95, // Dummy metric specific to SWELancer
+				timeEfficiency: 0.85, // Dummy metric specific to SWELancer
+			},
+		}
+	}
+}

evals/cli/src/adapters/types.ts

@@ -0,0 +1,31 @@
+/**
+ * Represents a task to be executed
+ */
+export interface Task {
+	id: string
+	name: string
+	description: string
+	workspacePath: string
+	setupCommands: string[]
+	verificationCommands: string[]
+	metadata: Record<string, any>
+}
+
+/**
+ * Result of verifying a task execution
+ */
+export interface VerificationResult {
+	success: boolean
+	metrics: Record<string, any>
+}
+
+/**
+ * Interface for benchmark adapters
+ */
+export interface BenchmarkAdapter {
+	name: string
+	setup(): Promise<void>
+	listTasks(): Promise<Task[]>
+	prepareTask(taskId: string): Promise<Task>
+	verifyResult(task: Task, result: any): Promise<VerificationResult>
+}

evals/cli/src/commands/report.ts

@@ -0,0 +1,237 @@
+import * as fs from "fs"
+import * as path from "path"
+import chalk from "chalk"
+import ora from "ora"
+import { ResultsDatabase } from "../db"
+import { generateMarkdownReport } from "../utils/markdown"
+
+interface ReportOptions {
+	format?: "json" | "markdown"
+	output?: string
+}
+
+/**
+ * Handler for the report command
+ * @param options Command options
+ */
+export async function reportHandler(options: ReportOptions): Promise<void> {
+	const format = options.format || "markdown"
+	const db = new ResultsDatabase()
+
+	try {
+		const spinner = ora("Generating report...").start()
+
+		// Get all runs
+		const runs = db.getRuns()
+
+		console.log(chalk.blue(`Found ${runs.length} evaluation runs`))
+
+		if (runs.length === 0) {
+			spinner.fail("No evaluation runs found")
+			return
+		}
+
+		// Generate summary report
+		const summary = {
+			runs: runs.length,
+			models: [...new Set(runs.map((run) => run.model))],
+			benchmarks: [...new Set(runs.map((run) => run.benchmark))],
+			tasks: 0,
+			successRate: 0,
+			averageTokens: 0,
+			averageCost: 0,
+			averageDuration: 0,
+			totalToolCalls: 0,
+			totalToolFailures: 0,
+			toolSuccessRate: 0,
+			toolUsage: {} as Record<string, { calls: number; failures: number }>,
+		}
+
+		let totalTasks = 0
+		let successfulTasks = 0
+		let totalTokens = 0
+		let totalCost = 0
+		let totalDuration = 0
+		let totalToolCalls = 0
+		let totalToolFailures = 0
+
+		for (const run of runs) {
+			const tasks = db.getRunTasks(run.id)
+			totalTasks += tasks.length
+
+			for (const task of tasks) {
+				if (task.success) {
+					successfulTasks++
+				}
+
+				const metrics = db.getTaskMetrics(task.id)
+
+				const tokensIn = metrics.find((m) => m.name === "tokensIn")?.value || 0
+				const tokensOut = metrics.find((m) => m.name === "tokensOut")?.value || 0
+				totalTokens += tokensIn + tokensOut
+
+				totalCost += metrics.find((m) => m.name === "cost")?.value || 0
+				totalDuration += metrics.find((m) => m.name === "duration")?.value || 0
+
+				// Collect tool call metrics
+				totalToolCalls += task.total_tool_calls || 0
+				totalToolFailures += task.total_tool_failures || 0
+
+				// Get detailed tool usage
+				const toolCalls = db.getTaskToolCalls(task.id)
+
+				for (const toolCall of toolCalls) {
+					if (!summary.toolUsage[toolCall.tool_name]) {
+						summary.toolUsage[toolCall.tool_name] = {
+							calls: 0,
+							failures: 0,
+						}
+					}
+
+					summary.toolUsage[toolCall.tool_name].calls += toolCall.call_count
+					summary.toolUsage[toolCall.tool_name].failures += toolCall.failure_count
+				}
+			}
+		}
+
+		// Calculate tool success rate
+		summary.totalToolCalls = totalToolCalls
+		summary.totalToolFailures = totalToolFailures
+		summary.toolSuccessRate = totalToolCalls > 0 ? 1 - totalToolFailures / totalToolCalls : 1.0
+
+		summary.tasks = totalTasks
+		summary.successRate = totalTasks > 0 ? successfulTasks / totalTasks : 0
+		summary.averageTokens = totalTasks > 0 ? totalTokens / totalTasks : 0
+		summary.averageCost = totalTasks > 0 ? totalCost / totalTasks : 0
+		summary.averageDuration = totalTasks > 0 ? totalDuration / totalTasks : 0
+
+		// Generate benchmark-specific reports
+		const benchmarkReports: Record<string, any> = {}
+
+		for (const benchmark of summary.benchmarks) {
+			const benchmarkRuns = runs.filter((run) => run.benchmark === benchmark)
+			const benchmarkSummary = {
+				runs: benchmarkRuns.length,
+				models: [...new Set(benchmarkRuns.map((run) => run.model))],
+				tasks: 0,
+				successRate: 0,
+				averageTokens: 0,
+				averageCost: 0,
+				averageDuration: 0,
+			}
+
+			let benchmarkTasks = 0
+			let benchmarkSuccessfulTasks = 0
+			let benchmarkTotalTokens = 0
+			let benchmarkTotalCost = 0
+			let benchmarkTotalDuration = 0
+
+			for (const run of benchmarkRuns) {
+				const tasks = db.getRunTasks(run.id)
+				benchmarkTasks += tasks.length
+
+				for (const task of tasks) {
+					if (task.success) {
+						benchmarkSuccessfulTasks++
+					}
+
+					const metrics = db.getTaskMetrics(task.id)
+
+					const tokensIn = metrics.find((m) => m.name === "tokensIn")?.value || 0
+					const tokensOut = metrics.find((m) => m.name === "tokensOut")?.value || 0
+					benchmarkTotalTokens += tokensIn + tokensOut
+
+					benchmarkTotalCost += metrics.find((m) => m.name === "cost")?.value || 0
+					benchmarkTotalDuration += metrics.find((m) => m.name === "duration")?.value || 0
+				}
+			}
+
+			benchmarkSummary.tasks = benchmarkTasks
+			benchmarkSummary.successRate = benchmarkTasks > 0 ? benchmarkSuccessfulTasks / benchmarkTasks : 0
+			benchmarkSummary.averageTokens = benchmarkTasks > 0 ? benchmarkTotalTokens / benchmarkTasks : 0
+			benchmarkSummary.averageCost = benchmarkTasks > 0 ? benchmarkTotalCost / benchmarkTasks : 0
+			benchmarkSummary.averageDuration = benchmarkTasks > 0 ? benchmarkTotalDuration / benchmarkTasks : 0
+
+			benchmarkReports[benchmark] = benchmarkSummary
+		}
+
+		// Generate model-specific reports
+		const modelReports: Record<string, any> = {}
+
+		for (const model of summary.models) {
+			const modelRuns = runs.filter((run) => run.model === model)
+			const modelSummary = {
+				runs: modelRuns.length,
+				benchmarks: [...new Set(modelRuns.map((run) => run.benchmark))],
+				tasks: 0,
+				successRate: 0,
+				averageTokens: 0,
+				averageCost: 0,
+				averageDuration: 0,
+			}
+
+			let modelTasks = 0
+			let modelSuccessfulTasks = 0
+			let modelTotalTokens = 0
+			let modelTotalCost = 0
+			let modelTotalDuration = 0
+
+			for (const run of modelRuns) {
+				const tasks = db.getRunTasks(run.id)
+				modelTasks += tasks.length
+
+				for (const task of tasks) {
+					if (task.success) {
+						modelSuccessfulTasks++
+					}
+
+					const metrics = db.getTaskMetrics(task.id)
+
+					const tokensIn = metrics.find((m) => m.name === "tokensIn")?.value || 0
+					const tokensOut = metrics.find((m) => m.name === "tokensOut")?.value || 0
+					modelTotalTokens += tokensIn + tokensOut
+
+					modelTotalCost += metrics.find((m) => m.name === "cost")?.value || 0
+					modelTotalDuration += metrics.find((m) => m.name === "duration")?.value || 0
+				}
+			}
+
+			modelSummary.tasks = modelTasks
+			modelSummary.successRate = modelTasks > 0 ? modelSuccessfulTasks / modelTasks : 0
+			modelSummary.averageTokens = modelTasks > 0 ? modelTotalTokens / modelTasks : 0
+			modelSummary.averageCost = modelTasks > 0 ? modelTotalCost / modelTasks : 0
+			modelSummary.averageDuration = modelTasks > 0 ? modelTotalDuration / modelTasks : 0
+
+			modelReports[model] = modelSummary
+		}
+
+		// Save reports
+		const reportDir = path.join(path.resolve(__dirname, "../../../"), "results", "reports")
+		fs.mkdirSync(reportDir, { recursive: true })
+
+		const timestamp = new Date().toISOString().replace(/:/g, "-")
+
+		if (format === "json") {
+			// Save JSON reports
+			fs.writeFileSync(path.join(reportDir, `summary-${timestamp}.json`), JSON.stringify(summary, null, 2))
+
+			fs.writeFileSync(path.join(reportDir, `benchmarks-${timestamp}.json`), JSON.stringify(benchmarkReports, null, 2))
+
+			fs.writeFileSync(path.join(reportDir, `models-${timestamp}.json`), JSON.stringify(modelReports, null, 2))
+
+			spinner.succeed(`JSON reports generated in ${reportDir}`)
+		} else {
+			// Generate markdown report
+			const outputPath = options.output || path.join(reportDir, `report-${timestamp}.md`)
+
+			generateMarkdownReport(summary, benchmarkReports, modelReports, outputPath)
+
+			spinner.succeed(`Markdown report generated at ${outputPath}`)
+		}
+	} catch (error: any) {
+		console.error(chalk.red(`Error generating report: ${error.message}`))
+		console.error(error.stack)
+	} finally {
+		db.close()
+	}
+}

evals/cli/src/commands/run.ts

@@ -0,0 +1,133 @@
+import * as path from "path"
+import { v4 as uuidv4 } from "uuid"
+import chalk from "chalk"
+import ora from "ora"
+import { getAdapter } from "../adapters"
+import { ResultsDatabase } from "../db"
+import { spawnVSCode, cleanupVSCode } from "../utils/vscode"
+import { sendTaskToServer } from "../utils/task"
+import { storeTaskResult } from "../utils/results"
+
+interface RunOptions {
+	benchmark?: string
+	model: string
+	count?: number
+	apiKey?: string
+}
+
+/**
+ * Handler for the run command
+ * @param options Command options
+ */
+export async function runHandler(options: RunOptions): Promise<void> {
+	// Determine which benchmarks to run
+	const benchmarks = options.benchmark ? [options.benchmark] : ["exercism"] // Default to exercism for now
+	const model = options.model
+	const count = options.count || Infinity
+
+	console.log(chalk.blue(`Running evaluations for model: ${model}`))
+	console.log(chalk.blue(`Benchmarks: ${benchmarks.join(", ")}`))
+
+	// Create a run for each benchmark
+	for (const benchmark of benchmarks) {
+		const runId = uuidv4()
+		const db = new ResultsDatabase()
+
+		console.log(chalk.green(`\nStarting run for benchmark: ${benchmark}`))
+
+		// Create run in database
+		db.createRun(runId, model, benchmark)
+
+		// Get adapter for this benchmark
+		try {
+			const adapter = getAdapter(benchmark)
+
+			// List tasks
+			const spinner = ora("Listing tasks...").start()
+			const tasks = await adapter.listTasks()
+			spinner.succeed(`Found ${tasks.length} tasks for ${benchmark}`)
+
+			// Limit number of tasks if specified
+			const tasksToRun = tasks.slice(0, count)
+
+			console.log(chalk.blue(`Running ${tasksToRun.length} tasks...`))
+
+			// Run each task
+			for (let i = 0; i < tasksToRun.length; i++) {
+				const task = tasksToRun[i]
+
+				console.log(chalk.cyan(`\nTask ${i + 1}/${tasksToRun.length}: ${task.name}`))
+
+				// Prepare task
+				const prepareSpinner = ora("Preparing task...").start()
+				const preparedTask = await adapter.prepareTask(task.id)
+				prepareSpinner.succeed("Task prepared")
+
+				// Spawn VSCode
+				console.log("Spawning VSCode...")
+				await spawnVSCode(preparedTask.workspacePath)
+
+				// Send task to server
+				const sendSpinner = ora("Sending task to server...").start()
+				try {
+					const result = await sendTaskToServer(preparedTask.description, options.apiKey)
+					sendSpinner.succeed("Task completed")
+
+					// Verify result
+					const verifySpinner = ora("Verifying result...").start()
+					const verification = await adapter.verifyResult(preparedTask, result)
+
+					if (verification.success) {
+						verifySpinner.succeed(
+							`Verification successful: ${verification.metrics.testsPassed}/${verification.metrics.testsTotal} tests passed`,
+						)
+					} else {
+						verifySpinner.fail(
+							`Verification failed: ${verification.metrics.testsPassed}/${verification.metrics.testsTotal} tests passed`,
+						)
+					}
+
+					// Store result
+					const storeSpinner = ora("Storing result...").start()
+					await storeTaskResult(runId, preparedTask, result, verification)
+					storeSpinner.succeed("Result stored")
+
+					console.log(chalk.green(`Task completed. Success: ${verification.success}`))
+
+					// Clean up VS Code and temporary files
+					const cleanupSpinner = ora("Cleaning up...").start()
+					try {
+						await cleanupVSCode(preparedTask.workspacePath)
+						cleanupSpinner.succeed("Cleanup completed")
+					} catch (cleanupError: any) {
+						cleanupSpinner.fail(`Cleanup failed: ${cleanupError.message}`)
+						console.error(chalk.yellow(cleanupError.stack))
+					}
+				} catch (error: any) {
+					sendSpinner.fail(`Task failed: ${error.message}`)
+					console.error(chalk.red(error.stack))
+
+					// Clean up VS Code and temporary files even if the task failed
+					const cleanupSpinner = ora("Cleaning up...").start()
+					try {
+						await cleanupVSCode(preparedTask.workspacePath)
+						cleanupSpinner.succeed("Cleanup completed")
+					} catch (cleanupError: any) {
+						cleanupSpinner.fail(`Cleanup failed: ${cleanupError.message}`)
+						console.error(chalk.yellow(cleanupError.stack))
+					}
+				}
+			}
+
+			// Mark run as complete
+			db.completeRun(runId)
+
+			console.log(chalk.green(`\nRun complete for benchmark: ${benchmark}`))
+		} catch (error: any) {
+			console.error(chalk.red(`Error running benchmark ${benchmark}: ${error.message}`))
+			console.error(error.stack)
+		}
+	}
+
+	console.log(chalk.green("\nAll evaluations complete"))
+}

evals/cli/src/commands/setup.ts

@@ -0,0 +1,72 @@
+import * as path from "path"
+import * as fs from "fs"
+import execa from "execa"
+import chalk from "chalk"
+import ora from "ora"
+import { getAllAdapters } from "../adapters/index"
+import { BenchmarkAdapter } from "../adapters/types"
+
+interface SetupOptions {
+	benchmarks: string
+}
+
+/**
+ * Handler for the setup command
+ * @param options Command options
+ */
+export async function setupHandler(options: SetupOptions): Promise<void> {
+	const benchmarks = options.benchmarks.split(",")
+
+	console.log(chalk.blue(`Setting up benchmarks: ${benchmarks.join(", ")}`))
+
+	// Create directories
+	const evalsDir = path.resolve(__dirname, "../../../")
+	const reposDir = path.join(evalsDir, "repositories")
+	const resultsDir = path.join(evalsDir, "results")
+
+	const spinner = ora("Creating directory structure").start()
+
+	try {
+		fs.mkdirSync(reposDir, { recursive: true })
+		fs.mkdirSync(resultsDir, { recursive: true })
+		fs.mkdirSync(path.join(resultsDir, "runs"), { recursive: true })
+		fs.mkdirSync(path.join(resultsDir, "reports"), { recursive: true })
+		spinner.succeed("Directory structure created")
+	} catch (error) {
+		spinner.fail(`Failed to create directory structure: ${(error as Error).message}`)
+		throw error
+	}
+
+	// Set up each benchmark
+	try {
+		const adapters = getAllAdapters().filter((adapter: BenchmarkAdapter) => benchmarks.includes(adapter.name))
+
+		if (adapters.length === 0) {
+			console.warn(chalk.yellow("No valid benchmarks specified. Available benchmarks:"))
+			console.warn(
+				chalk.yellow(
+					getAllAdapters()
+						.map((a: BenchmarkAdapter) => a.name)
+						.join(", "),
+				),
+			)
+			return
+		}
+
+		for (const adapter of adapters) {
+			const setupSpinner = ora(`Setting up ${adapter.name}...`).start()
+			try {
+				await adapter.setup()
+				setupSpinner.succeed(`${adapter.name} setup complete`)
+			} catch (error) {
+				setupSpinner.fail(`Failed to set up ${adapter.name}: ${(error as Error).message}`)
+				throw error
+			}
+		}
+
+		console.log(chalk.green("Setup complete"))
+	} catch (error) {
+		console.error(chalk.red(`Setup failed: ${(error as Error).message}`))
+		throw error
+	}
+}

evals/cli/src/db/index.ts

@@ -0,0 +1,211 @@
+import * as path from "path"
+import * as fs from "fs"
+import Database from "better-sqlite3"
+import { SCHEMA } from "./schema"
+
+const EVALS_DIR = path.resolve(__dirname, "../../../")
+
+/**
+ * Database class for storing evaluation results
+ */
+export class ResultsDatabase {
+	db: Database.Database
+
+	constructor() {
+		// Ensure results directory exists
+		const resultsDir = path.join(EVALS_DIR, "results")
+		fs.mkdirSync(resultsDir, { recursive: true })
+
+		// Create database file
+		const dbPath = path.join(resultsDir, "evals.db")
+		this.db = new Database(dbPath)
+
+		// Initialize schema
+		this.initSchema()
+	}
+
+	/**
+	 * Initialize the database schema
+	 */
+	private initSchema(): void {
+		this.db.exec(SCHEMA)
+	}
+
+	/**
+	 * Create a new evaluation run
+	 * @param id Run ID
+	 * @param model Model name
+	 * @param benchmark Benchmark name
+	 */
+	createRun(id: string, model: string, benchmark: string): void {
+		const stmt = this.db.prepare(`
+      INSERT INTO runs (id, timestamp, model, benchmark)
+      VALUES (?, ?, ?, ?)
+    `)
+
+		stmt.run(id, Date.now(), model, benchmark)
+	}
+
+	/**
+	 * Mark a run as completed
+	 * @param id Run ID
+	 */
+	completeRun(id: string): void {
+		const stmt = this.db.prepare(`
+      UPDATE runs SET completed = 1 WHERE id = ?
+    `)
+
+		stmt.run(id)
+	}
+
+	/**
+	 * Create a new task
+	 * @param id Task ID
+	 * @param runId Run ID
+	 * @param taskId Original task ID
+	 */
+	createTask(id: string, runId: string, taskId: string): void {
+		const stmt = this.db.prepare(`
+      INSERT INTO tasks (id, run_id, task_id, timestamp)
+      VALUES (?, ?, ?, ?)
+    `)
+
+		stmt.run(id, runId, taskId, Date.now())
+	}
+
+	/**
+	 * Mark a task as completed
+	 * @param id Task ID
+	 * @param success Whether the task was successful
+	 * @param toolCalls Total tool calls
+	 * @param toolFailures Total tool failures
+	 */
+	completeTask(id: string, success: boolean, toolCalls: number = 0, toolFailures: number = 0): void {
+		const stmt = this.db.prepare(`
+      UPDATE tasks 
+      SET success = ?, total_tool_calls = ?, total_tool_failures = ? 
+      WHERE id = ?
+    `)
+
+		stmt.run(success ? 1 : 0, toolCalls, toolFailures, id)
+	}
+
+	/**
+	 * Add a metric to a task
+	 * @param taskId Task ID
+	 * @param name Metric name
+	 * @param value Metric value
+	 */
+	addMetric(taskId: string, name: string, value: number): void {
+		const stmt = this.db.prepare(`
+      INSERT INTO metrics (task_id, name, value)
+      VALUES (?, ?, ?)
+    `)
+
+		stmt.run(taskId, name, value)
+	}
+
+	/**
+	 * Add a tool call record
+	 * @param taskId Task ID
+	 * @param toolName Tool name
+	 * @param callCount Number of calls
+	 * @param failureCount Number of failures
+	 */
+	addToolCall(taskId: string, toolName: string, callCount: number, failureCount: number): void {
+		const stmt = this.db.prepare(`
+      INSERT INTO tool_calls (task_id, tool_name, call_count, failure_count)
+      VALUES (?, ?, ?, ?)
+    `)
+
+		stmt.run(taskId, toolName, callCount, failureCount)
+	}
+
+	/**
+	 * Add a file record
+	 * @param taskId Task ID
+	 * @param filePath File path
+	 * @param status File status (created, modified, deleted)
+	 */
+	addFile(taskId: string, filePath: string, status: "created" | "modified" | "deleted"): void {
+		const stmt = this.db.prepare(`
+      INSERT INTO files (task_id, path, status)
+      VALUES (?, ?, ?)
+    `)
+
+		stmt.run(taskId, filePath, status)
+	}
+
+	/**
+	 * Get all runs
+	 * @returns Array of runs
+	 */
+	getRuns(): any[] {
+		const stmt = this.db.prepare(`
+      SELECT * FROM runs ORDER BY timestamp DESC
+    `)
+
+		return stmt.all()
+	}
+
+	/**
+	 * Get all tasks for a run
+	 * @param runId Run ID
+	 * @returns Array of tasks
+	 */
+	getRunTasks(runId: string): any[] {
+		const stmt = this.db.prepare(`
+      SELECT * FROM tasks WHERE run_id = ? ORDER BY timestamp ASC
+    `)
+
+		return stmt.all(runId)
+	}
+
+	/**
+	 * Get all metrics for a task
+	 * @param taskId Task ID
+	 * @returns Array of metrics
+	 */
+	getTaskMetrics(taskId: string): any[] {
+		const stmt = this.db.prepare(`
+      SELECT name, value FROM metrics WHERE task_id = ?
+    `)
+
+		return stmt.all(taskId)
+	}
+
+	/**
+	 * Get all tool calls for a task
+	 * @param taskId Task ID
+	 * @returns Array of tool calls
+	 */
+	getTaskToolCalls(taskId: string): any[] {
+		const stmt = this.db.prepare(`
+      SELECT tool_name, call_count, failure_count 
+      FROM tool_calls 
+      WHERE task_id = ?
+    `)
+
+		return stmt.all(taskId)
+	}
+
+	/**
+	 * Get all files for a task
+	 * @param taskId Task ID
+	 * @returns Array of files
+	 */
+	getTaskFiles(taskId: string): any[] {
+		const stmt = this.db.prepare(`
+      SELECT path, status FROM files WHERE task_id = ?
+    `)
+
+		return stmt.all(taskId)
+	}
+
+	/**
+	 * Close the database connection
+	 */
+	close(): void {
+		this.db.close()
+	}
+}

evals/cli/src/db/schema.ts

@@ -0,0 +1,48 @@
+/**
+ * SQL schema for the evaluation database
+ */
+export const SCHEMA = `
+CREATE TABLE IF NOT EXISTS runs (
+  id TEXT PRIMARY KEY,
+  timestamp INTEGER NOT NULL,
+  model TEXT NOT NULL,
+  benchmark TEXT NOT NULL,
+  completed INTEGER NOT NULL DEFAULT 0
+);
+
+CREATE TABLE IF NOT EXISTS tasks (
+  id TEXT PRIMARY KEY,
+  run_id TEXT NOT NULL,
+  task_id TEXT NOT NULL,
+  timestamp INTEGER NOT NULL,
+  success INTEGER NOT NULL DEFAULT 0,
+  total_tool_calls INTEGER DEFAULT 0,
+  total_tool_failures INTEGER DEFAULT 0,
+  FOREIGN KEY (run_id) REFERENCES runs(id)
+);
+
+CREATE TABLE IF NOT EXISTS metrics (
+  id INTEGER PRIMARY KEY AUTOINCREMENT,
+  task_id TEXT NOT NULL,
+  name TEXT NOT NULL,
+  value REAL NOT NULL,
+  FOREIGN KEY (task_id) REFERENCES tasks(id)
+);
+
+CREATE TABLE IF NOT EXISTS tool_calls (
+  id INTEGER PRIMARY KEY AUTOINCREMENT,
+  task_id TEXT NOT NULL,
+  tool_name TEXT NOT NULL,
+  call_count INTEGER NOT NULL,
+  failure_count INTEGER NOT NULL,
+  FOREIGN KEY (task_id) REFERENCES tasks(id)
+);
+
+CREATE TABLE IF NOT EXISTS files (
+  id INTEGER PRIMARY KEY AUTOINCREMENT,
+  task_id TEXT NOT NULL,
+  path TEXT NOT NULL,
+  status TEXT NOT NULL,
+  FOREIGN KEY (task_id) REFERENCES tasks(id)
+);
+`

evals/cli/src/index.ts

@@ -0,0 +1,70 @@
+#!/usr/bin/env node
+import { Command } from "commander"
+import chalk from "chalk"
+import { setupHandler } from "./commands/setup"
+import { runHandler } from "./commands/run"
+import { reportHandler } from "./commands/report"
+
+// Create the CLI program
+const program = new Command()
+
+// Set up CLI metadata
+program.name("cline-eval").description("CLI tool for orchestrating Cline evaluations across multiple benchmarks").version("0.1.0")
+
+// Setup command
+program
+	.command("setup")
+	.description("Clone and set up benchmark repositories")
+	.option(
+		"-b, --benchmarks <benchmarks>",
+		"Comma-separated list of benchmarks to set up",
+		"exercism,swe-bench,swelancer,multi-swe",
+	)
+	.action(async (options) => {
+		try {
+			await setupHandler(options)
+		} catch (error) {
+			console.error(chalk.red(`Error during setup: ${error instanceof Error ? error.message : String(error)}`))
+			process.exit(1)
+		}
+	})
+
+// Run command
+program
+	.command("run")
+	.description("Run evaluations")
+	.option("-b, --benchmark <benchmark>", "Specific benchmark to run")
+	.option("-m, --model <model>", "Model to evaluate", "claude-3-opus-20240229")
+	.option("-c, --count <count>", "Number of tasks to run", parseInt)
+	.option("-k, --api-key <apiKey>", "Cline API key to use for evaluations")
+	.action(async (options) => {
+		try {
+			await runHandler(options)
+		} catch (error) {
+			console.error(chalk.red(`Error during run: ${error instanceof Error ? error.message : String(error)}`))
+			process.exit(1)
+		}
+	})
+
+// Report command
+program
+	.command("report")
+	.description("Generate reports")
+	.option("-f, --format <format>", "Report format (json, markdown)", "markdown")
+	.option("-o, --output <path>", "Output path for the report")
+	.action(async (options) => {
+		try {
+			await reportHandler(options)
+		} catch (error) {
+			console.error(chalk.red(`Error generating report: ${error instanceof Error ? error.message : String(error)}`))
+			process.exit(1)
+		}
+	})
+
+// Parse command line arguments
+program.parse(process.argv)
+
+// If no arguments provided, show help
+if (process.argv.length === 2) {
+	program.help()
+}

evals/cli/src/utils/extensions.ts

@@ -0,0 +1,131 @@
+import execa from "execa"
+import * as fs from "fs"
+import * as path from "path"
+import * as os from "os"
+
+/**
+ * List of VSCode extensions to install for evaluation environments
+ * These extensions provide language support and other useful features
+ */
+export const REQUIRED_EXTENSIONS = [
+	"golang.go", // Go language support
+	"dbaeumer.vscode-eslint", // ESLint support
+	"redhat.java", // Java support
+	"ms-python.python", // Python support
+	"rust-lang.rust-analyzer", // Rust support
+	"ms-vscode.cpptools", // C/C++ support
+]
+
+/**
+ * Install required VSCode extensions in the specified extensions directory
+ * @param extensionsDir The directory where extensions should be installed
+ * @returns Promise that resolves when all extensions are installed
+ */
+export async function installRequiredExtensions(extensionsDir: string): Promise<void> {
+	console.log("Installing required VSCode extensions...")
+
+	// Create the extensions directory if it doesn't exist
+	if (!fs.existsSync(extensionsDir)) {
+		fs.mkdirSync(extensionsDir, { recursive: true })
+	}
+
+	// Install each extension
+	for (const extension of REQUIRED_EXTENSIONS) {
+		try {
+			console.log(`Installing extension: ${extension}...`)
+			await execa("code", ["--extensions-dir", extensionsDir, "--install-extension", extension, "--force"])
+			console.log(`✅ Extension ${extension} installed successfully`)
+		} catch (error: any) {
+			console.warn(`⚠️ Failed to install extension ${extension}: ${error.message}`)
+			// Continue with other extensions even if one fails
+		}
+	}
+
+	console.log("✅ All required extensions installed")
+}
+
+/**
+ * Check if a VSCode extension is installed in the specified directory
+ * @param extensionsDir The directory to check for installed extensions
+ * @param extensionId The ID of the extension to check
+ * @returns True if the extension is installed, false otherwise
+ */
+export function isExtensionInstalled(extensionsDir: string, extensionId: string): boolean {
+	// Extensions are installed in directories named publisher.name-version
+	// We need to check if any directory starts with the extensionId
+	const extensionPrefix = extensionId.toLowerCase() + "-"
+
+	try {
+		const files = fs.readdirSync(extensionsDir)
+		return files.some((file) => {
+			const lowerCaseFile = file.toLowerCase()
+			return lowerCaseFile === extensionId.toLowerCase() || lowerCaseFile.startsWith(extensionPrefix)
+		})
+	} catch (error) {
+		return false
+	}
+}
+
+/**
+ * Get the path to the VSCode settings file in the specified user data directory
+ * @param userDataDir The VSCode user data directory
+ * @returns The path to the settings.json file
+ */
+export function getSettingsPath(userDataDir: string): string {
+	const settingsDir = path.join(userDataDir, "User")
+	fs.mkdirSync(settingsDir, { recursive: true })
+	return path.join(settingsDir, "settings.json")
+}
+
+/**
+ * Configure extension settings in the VSCode user data directory
+ * @param userDataDir The VSCode user data directory
+ */
+export function configureExtensionSettings(userDataDir: string): void {
+	const settingsPath = getSettingsPath(userDataDir)
+
+	// Read existing settings if they exist
+	let settings = {}
+	if (fs.existsSync(settingsPath)) {
+		try {
+			settings = JSON.parse(fs.readFileSync(settingsPath, "utf8"))
+		} catch (error) {
+			console.warn(`Error reading settings file: ${error}`)
+		}
+	}
+
+	// Add or update extension-specific settings
+	const updatedSettings = {
+		...settings,
+		// Go extension settings
+		"go.toolsManagement.autoUpdate": false,
+		"go.survey.prompt": false,
+
+		// ESLint settings
+		"eslint.enable": true,
+		"eslint.run": "onSave",
+
+		// Java settings
+		"java.configuration.checkProjectSettingsExclusions": false,
+		"java.configure.checkForOutdatedExtensions": false,
+		"java.help.firstView": false,
+
+		// Python settings
+		"python.experiments.enabled": false,
+		"python.showStartPage": false,
+
+		// Rust settings
+		"rust-analyzer.checkOnSave.command": "check",
+
+		// C/C++ settings
+		"C_Cpp.intelliSenseEngine": "default",
+
+		// General extension settings
+		"extensions.autoUpdate": false,
+		"extensions.ignoreRecommendations": true,
+	}
+
+	// Write updated settings
+	fs.writeFileSync(settingsPath, JSON.stringify(updatedSettings, null, 2))
+	console.log("✅ Extension settings configured")
+}

evals/cli/src/utils/markdown.ts

@@ -0,0 +1,109 @@
+import * as fs from "fs"
+import * as path from "path"
+
+/**
+ * Generate a markdown report from evaluation results
+ * @param summary Overall summary
+ * @param benchmarkReports Benchmark-specific reports
+ * @param modelReports Model-specific reports
+ * @param outputPath Output file path
+ */
+export function generateMarkdownReport(
+	summary: any,
+	benchmarkReports: Record<string, any>,
+	modelReports: Record<string, any>,
+	outputPath: string,
+): void {
+	let markdown = `# Cline Evaluation Report\n\n`
+
+	// Generate summary section
+	markdown += `## Summary\n\n`
+	markdown += `- **Total Runs:** ${summary.runs}\n`
+	markdown += `- **Models:** ${summary.models.join(", ")}\n`
+	markdown += `- **Benchmarks:** ${summary.benchmarks.join(", ")}\n`
+	markdown += `- **Total Tasks:** ${summary.tasks}\n`
+	markdown += `- **Success Rate:** ${(summary.successRate * 100).toFixed(2)}%\n`
+	markdown += `- **Average Tokens:** ${Math.round(summary.averageTokens)}\n`
+	markdown += `- **Average Cost:** $${summary.averageCost.toFixed(4)}\n`
+	markdown += `- **Average Duration:** ${(summary.averageDuration / 1000).toFixed(2)}s\n`
+	markdown += `- **Total Tool Calls:** ${summary.totalToolCalls}\n`
+	markdown += `- **Tool Success Rate:** ${(summary.toolSuccessRate * 100).toFixed(2)}%\n\n`
+
+	// Generate tool usage section
+	markdown += `## Tool Usage\n\n`
+	markdown += `| Tool | Calls | Failures | Success Rate |\n`
+	markdown += `| ---- | ----- | -------- | ------------ |\n`
+
+	for (const [toolName, metrics] of Object.entries(summary.toolUsage)) {
+		const calls = (metrics as any).calls
+		const failures = (metrics as any).failures
+		const successRate = calls > 0 ? (1 - failures / calls) * 100 : 100
+
+		markdown += `| ${toolName} | ${calls} | ${failures} | ${successRate.toFixed(2)}% |\n`
+	}
+
+	// Generate benchmark results section
+	markdown += `\n## Benchmark Results\n\n`
+
+	for (const [benchmark, report] of Object.entries(benchmarkReports)) {
+		markdown += `### ${benchmark}\n\n`
+		markdown += `- **Runs:** ${report.runs}\n`
+		markdown += `- **Models:** ${report.models.join(", ")}\n`
+		markdown += `- **Tasks:** ${report.tasks}\n`
+		markdown += `- **Success Rate:** ${(report.successRate * 100).toFixed(2)}%\n`
+		markdown += `- **Average Tokens:** ${Math.round(report.averageTokens)}\n`
+		markdown += `- **Average Cost:** $${report.averageCost.toFixed(4)}\n`
+		markdown += `- **Average Duration:** ${(report.averageDuration / 1000).toFixed(2)}s\n\n`
+	}
+
+	// Generate model results section
+	markdown += `## Model Results\n\n`
+
+	for (const [model, report] of Object.entries(modelReports)) {
+		markdown += `### ${model}\n\n`
+		markdown += `- **Runs:** ${report.runs}\n`
+		markdown += `- **Benchmarks:** ${report.benchmarks.join(", ")}\n`
+		markdown += `- **Tasks:** ${report.tasks}\n`
+		markdown += `- **Success Rate:** ${(report.successRate * 100).toFixed(2)}%\n`
+		markdown += `- **Average Tokens:** ${Math.round(report.averageTokens)}\n`
+		markdown += `- **Average Cost:** $${report.averageCost.toFixed(4)}\n`
+		markdown += `- **Average Duration:** ${(report.averageDuration / 1000).toFixed(2)}s\n\n`
+	}
+
+	// Add charts using Mermaid
+	markdown += `## Charts\n\n`
+
+	// Success rate by benchmark chart
+	markdown += `### Success Rate by Benchmark\n\n`
+	markdown += "```mermaid\n"
+	markdown += "graph TD\n"
+	markdown += "  title[Success Rate by Benchmark]\n"
+	markdown += "  style title fill:none,stroke:none\n\n"
+
+	for (const [benchmark, report] of Object.entries(benchmarkReports)) {
+		const successRate = (report.successRate * 100).toFixed(2)
+		markdown += `  ${benchmark}[${benchmark}: ${successRate}%]\n`
+	}
+
+	markdown += "```\n\n"
+
+	// Success rate by model chart
+	markdown += `### Success Rate by Model\n\n`
+	markdown += "```mermaid\n"
+	markdown += "graph TD\n"
+	markdown += "  title[Success Rate by Model]\n"
+	markdown += "  style title fill:none,stroke:none\n\n"
+
+	for (const [model, report] of Object.entries(modelReports)) {
+		const successRate = (report.successRate * 100).toFixed(2)
+		markdown += `  ${model.replace(/[-\.]/g, "_")}[${model}: ${successRate}%]\n`
+	}
+
+	markdown += "```\n\n"
+
+	// Add timestamp
+	markdown += `\n\n---\n\nReport generated on ${new Date().toISOString()}\n`
+
+	// Write markdown to file
+	fs.writeFileSync(outputPath, markdown)
+}

evals/cli/src/utils/results.ts

@@ -0,0 +1,79 @@
+import { v4 as uuidv4 } from "uuid"
+import { ResultsDatabase } from "../db"
+import { Task } from "../adapters/types"
+
+/**
+ * Store task result in the database
+ * @param runId The run ID
+ * @param task The task that was executed
+ * @param result The result from the test server
+ * @param verification The verification result
+ */
+export async function storeTaskResult(runId: string, task: Task, result: any, verification: any): Promise<void> {
+	const db = new ResultsDatabase()
+	const taskId = uuidv4()
+
+	try {
+		// Extract metrics from the result
+		const { metrics } = result
+		const totalToolCalls = metrics?.totalToolCalls || 0
+		const totalToolFailures = metrics?.totalToolFailures || 0
+
+		// Create task with tool metrics
+		db.createTask(taskId, runId, task.id)
+		db.completeTask(taskId, verification.success, totalToolCalls, totalToolFailures)
+
+		// Store metrics
+		if (metrics) {
+			// Store token metrics
+			if (metrics.tokensIn) db.addMetric(taskId, "tokensIn", metrics.tokensIn)
+			if (metrics.tokensOut) db.addMetric(taskId, "tokensOut", metrics.tokensOut)
+			if (metrics.cost) db.addMetric(taskId, "cost", metrics.cost)
+			if (metrics.duration) db.addMetric(taskId, "duration", metrics.duration)
+
+			// Store tool call metrics
+			if (metrics.toolCalls) {
+				for (const [toolName, callCount] of Object.entries(metrics.toolCalls)) {
+					const failureCount = metrics.toolFailures?.[toolName] || 0
+					db.addToolCall(taskId, toolName, callCount as number, failureCount)
+				}
+			}
+		}
+
+		// Store verification metrics
+		if (verification.metrics) {
+			for (const [key, value] of Object.entries(verification.metrics)) {
+				if (typeof value === "number") {
+					db.addMetric(taskId, key, value)
+				}
+			}
+		}
+
+		// Store file changes
+		if (result.files) {
+			// Store created files
+			if (result.files.created) {
+				for (const file of result.files.created) {
+					db.addFile(taskId, file, "created")
+				}
+			}
+
+			// Store modified files
+			if (result.files.modified) {
+				for (const file of result.files.modified) {
+					db.addFile(taskId, file, "modified")
+				}
+			}
+
+			// Store deleted files
+			if (result.files.deleted) {
+				for (const file of result.files.deleted) {
+					db.addFile(taskId, file, "deleted")
+				}
+			}
+		}
+	} finally {
+		// Close the database connection
+		db.close()
+	}
+}

evals/cli/src/utils/task.ts

@@ -0,0 +1,52 @@
+import fetch from "node-fetch"
+import chalk from "chalk"
+
+/**
+ * Send a task to the Cline test server
+ * @param task The task description to send
+ * @param apiKey Optional Cline API key to use for the task
+ * @returns The result of the task execution
+ */
+export async function sendTaskToServer(task: string, apiKey?: string): Promise<any> {
+	const SERVER_URL = "http://localhost:9876/task"
+
+	try {
+		console.log(chalk.blue(`Sending task to server: ${task.substring(0, 100)}${task.length > 100 ? "..." : ""}`))
+
+		const response = await fetch(SERVER_URL, {
+			method: "POST",
+			headers: {
+				"Content-Type": "application/json",
+			},
+			body: JSON.stringify({
+				task,
+				apiKey,
+			}),
+		})
+
+		if (!response.ok) {
+			const errorText = await response.text()
+			throw new Error(`Server responded with status ${response.status}: ${errorText}`)
+		}
+
+		const result = await response.json()
+
+		if (!result.success) {
+			throw new Error(`Task execution failed: ${result.error || "Unknown error"}`)
+		}
+
+		if (result.timeout) {
+			throw new Error("Task execution timed out")
+		}
+
+		return result
+	} catch (error: any) {
+		if (error.code === "ECONNREFUSED") {
+			throw new Error(
+				"Could not connect to the test server. Make sure VSCode is running with the Cline extension and the test server is active.",
+			)
+		}
+
+		throw error
+	}
+}

evals/cli/src/utils/vscode.ts

@@ -0,0 +1,570 @@
+import execa from "execa"
+import * as path from "path"
+import * as fs from "fs"
+import fetch from "node-fetch"
+import * as os from "os"
+import * as child_process from "child_process"
+import { installRequiredExtensions, configureExtensionSettings } from "./extensions"
+
+// Store temporary directories for cleanup
+interface VSCodeResources {
+	tempUserDataDir: string
+	tempExtensionsDir: string
+	vscodePid?: number
+}
+
+// Global map to track resources for each workspace
+const workspaceResources = new Map<string, VSCodeResources>()
+
+/**
+ * Spawn a VSCode instance with the Cline extension
+ * @param workspacePath The workspace path to open
+ * @param vsixPath Optional path to a VSIX file to install
+ * @returns The resources created for this VS Code instance
+ */
+export async function spawnVSCode(workspacePath: string, vsixPath?: string): Promise<VSCodeResources> {
+	// Ensure the workspace path exists
+	if (!fs.existsSync(workspacePath)) {
+		throw new Error(`Workspace path does not exist: ${workspacePath}`)
+	}
+
+	// If no VSIX path is provided, build one with IS_TEST=true
+	if (!vsixPath) {
+		try {
+			// Build the VSIX with IS_TEST=true
+			console.log("Building test VSIX...")
+			const clineRoot = path.resolve(process.cwd(), "..", "..")
+			await execa("npx", ["vsce", "package"], {
+				cwd: clineRoot,
+				env: {
+					IS_TEST: "true",
+				},
+				stdio: "inherit",
+			})
+
+			// Find the generated VSIX file
+			const files = fs.readdirSync(clineRoot)
+			const vsixFile = files.find((file) => file.endsWith(".vsix"))
+			if (vsixFile) {
+				vsixPath = path.join(clineRoot, vsixFile)
+				console.log(`Using built VSIX: ${vsixPath}`)
+			} else {
+				console.warn("Could not find generated VSIX file")
+			}
+		} catch (error) {
+			console.warn("Failed to build test VSIX:", error)
+		}
+	}
+
+	// Create a temporary user data directory for this VS Code instance
+	const tempUserDataDir = path.join(os.tmpdir(), `vscode-cline-eval-${Date.now()}`)
+	fs.mkdirSync(tempUserDataDir, { recursive: true })
+	console.log(`Created temporary user data directory: ${tempUserDataDir}`)
+
+	// Create a temporary extensions directory to ensure no other extensions are loaded
+	const tempExtensionsDir = path.join(os.tmpdir(), `vscode-cline-eval-ext-${Date.now()}`)
+	fs.mkdirSync(tempExtensionsDir, { recursive: true })
+	console.log(`Created temporary extensions directory: ${tempExtensionsDir}`)
+
+	// Create settings.json in the temporary user data directory to disable workspace trust
+	// and configure Cline to auto-open on startup
+	const settingsDir = path.join(tempUserDataDir, "User")
+	fs.mkdirSync(settingsDir, { recursive: true })
+	const settingsPath = path.join(settingsDir, "settings.json")
+	const settings = {
+		// Disable workspace trust
+		"security.workspace.trust.enabled": false,
+		"security.workspace.trust.startupPrompt": "never",
+		"security.workspace.trust.banner": "never",
+		"security.workspace.trust.emptyWindow": true,
+
+		// Configure startup behavior
+		"workbench.startupEditor": "none",
+
+		// Auto-open Cline on startup
+		"cline.autoOpenOnStartup": true,
+
+		// Show the activity bar and sidebar
+		"workbench.activityBar.visible": true,
+		"workbench.sideBar.visible": true,
+		"workbench.view.extension.saoudrizwan.claude-dev-ActivityBar.visible": true,
+		"workbench.view.alwaysShowHeaderActions": true,
+		"workbench.editor.openSideBySideDirection": "right",
+
+		// Disable GitLens from opening automatically
+		"gitlens.views.repositories.autoReveal": false,
+		"gitlens.views.fileHistory.autoReveal": false,
+		"gitlens.views.lineHistory.autoReveal": false,
+		"gitlens.views.compare.autoReveal": false,
+		"gitlens.views.search.autoReveal": false,
+		"gitlens.showWelcomeOnInstall": false,
+		"gitlens.showWhatsNewAfterUpgrades": false,
+
+		// Disable other extensions that might compete for startup focus
+		"extensions.autoUpdate": false,
+	}
+	fs.writeFileSync(settingsPath, JSON.stringify(settings, null, 2))
+	console.log(`Created settings.json to disable workspace trust and auto-open Cline`)
+
+	// Create keybindings.json to automatically open Cline on startup
+	const keybindingsPath = path.join(settingsDir, "keybindings.json")
+	const keybindings = [
+		{
+			key: "alt+c",
+			command: "workbench.view.extension.saoudrizwan.claude-dev-ActivityBar",
+			when: "viewContainer.workbench.view.extension.saoudrizwan.claude-dev-ActivityBar.enabled",
+		},
+		{
+			key: "alt+shift+c",
+			command: "cline.openInNewTab",
+			when: "viewContainer.workbench.view.extension.saoudrizwan.claude-dev-ActivityBar.enabled",
+		},
+	]
+	fs.writeFileSync(keybindingsPath, JSON.stringify(keybindings, null, 2))
+	console.log(`Created keybindings.json to help with Cline activation`)
+
+	// Build the command arguments with custom user data directory
+	const args = [
+		// Use a custom user data directory to isolate this instance
+		"--user-data-dir",
+		tempUserDataDir,
+		// Use a custom extensions directory to ensure only our extension is loaded
+		"--extensions-dir",
+		tempExtensionsDir,
+		// Disable workspace trust
+		"--disable-workspace-trust",
+		"-n",
+		workspacePath,
+		// Force the extension to be activated on startup
+		"--start-up-extension",
+		"saoudrizwan.claude-dev",
+		// Run a command on startup to open Cline
+		"--command",
+		"workbench.view.extension.saoudrizwan.claude-dev-ActivityBar",
+		// Additional flags to help with extension activation
+		"--disable-gpu=false",
+		"--max-memory=4096",
+	]
+
+	// Create a startup script to run commands after VS Code launches
+	const startupScriptPath = path.join(settingsDir, "startup.js")
+	const startupScript = `
+		// This script will be executed when VS Code starts
+		setTimeout(() => {
+			// Try to open Cline in the sidebar
+			require('vscode').commands.executeCommand('workbench.view.extension.saoudrizwan.claude-dev-ActivityBar');
+			
+			// Also try to open Cline in a tab as a fallback
+			setTimeout(() => {
+				require('vscode').commands.executeCommand('cline.openInNewTab');
+			}, 5000);
+		}, 5000);
+	`
+	fs.writeFileSync(startupScriptPath, startupScript)
+	console.log(`Created startup script to activate Cline`)
+
+	// If a VSIX is provided, install it
+	if (vsixPath) {
+		if (!fs.existsSync(vsixPath)) {
+			throw new Error(`VSIX file does not exist: ${vsixPath}`)
+		}
+		args.unshift("--install-extension", vsixPath)
+	}
+
+	// Install required extensions
+	console.log("Installing required VSCode extensions...")
+	await installRequiredExtensions(tempExtensionsDir)
+
+	// Configure extension settings
+	console.log("Configuring extension settings...")
+	configureExtensionSettings(tempUserDataDir)
+
+	// Execute the command
+	try {
+		// We don't need to install extensions globally anymore since we're using a custom user data directory
+		// The VSIX will be installed in the isolated environment if provided in the args
+
+		// Launch VS Code
+		console.log("Launching VS Code...")
+		await execa("code", args, {
+			stdio: "inherit",
+		})
+
+		// Wait longer for VSCode to initialize and extension to load
+		console.log("Waiting for VS Code to initialize...")
+		await new Promise((resolve) => setTimeout(resolve, 30000))
+
+		// Create a JavaScript file that will be loaded as a VS Code extension
+		const extensionDir = path.join(tempExtensionsDir, "cline-activator")
+		fs.mkdirSync(extensionDir, { recursive: true })
+
+		// Create package.json for the extension
+		const packageJsonPath = path.join(extensionDir, "package.json")
+		const packageJson = {
+			name: "cline-activator",
+			displayName: "Cline Activator",
+			description: "Activates Cline and starts the test server",
+			version: "0.0.1",
+			engines: {
+				vscode: "^1.60.0",
+			},
+			main: "./extension.js",
+			activationEvents: ["*"],
+			contributes: {
+				commands: [
+					{
+						command: "cline-activator.activate",
+						title: "Activate Cline",
+					},
+				],
+			},
+		}
+		fs.writeFileSync(packageJsonPath, JSON.stringify(packageJson, null, 2))
+
+		// Create extension.js
+		const extensionJsPath = path.join(extensionDir, "extension.js")
+		const extensionJs = `
+			const vscode = require('vscode');
+			
+			/**
+			 * @param {vscode.ExtensionContext} context
+			 */
+			function activate(context) {
+				console.log('Cline Activator is now active!');
+				
+				// Register the command to activate Cline
+				let disposable = vscode.commands.registerCommand('cline-activator.activate', async function () {
+					try {
+						// Make sure the Cline extension is activated
+						const extension = vscode.extensions.getExtension('saoudrizwan.claude-dev');
+						if (!extension) {
+							console.error('Cline extension not found');
+							return;
+						}
+						
+						if (!extension.isActive) {
+							console.log('Activating Cline extension...');
+							await extension.activate();
+						}
+						
+						// Show the Cline sidebar
+						console.log('Opening Cline sidebar...');
+						await vscode.commands.executeCommand('workbench.view.extension.saoudrizwan.claude-dev-ActivityBar');
+						
+						// Wait a moment for the sidebar to initialize
+						await new Promise(resolve => setTimeout(resolve, 2000));
+						
+						// Also open Cline in a tab as a fallback
+						console.log('Opening Cline in a tab...');
+						await vscode.commands.executeCommand('cline.openInNewTab');
+						
+						// Wait a moment for the tab to initialize
+						await new Promise(resolve => setTimeout(resolve, 2000));
+						
+						// Create the test server if it doesn't exist
+						console.log('Creating test server...');
+						
+						// Get the visible webview instance
+						const clineRootPath = '${path.resolve(process.cwd(), "..", "..")}';
+						const visibleWebview = require(path.join(clineRootPath, 'src', 'core', 'webview')).WebviewProvider.getVisibleInstance();
+						if (visibleWebview) {
+							require(path.join(clineRootPath, 'src', 'services', 'test', 'TestServer')).createTestServer(visibleWebview);
+							console.log('Test server created successfully');
+						} else {
+							console.error('No visible webview instance found');
+						}
+					} catch (error) {
+						console.error('Error activating Cline:', error);
+					}
+				});
+				
+				context.subscriptions.push(disposable);
+				
+				// Automatically run the command after a delay
+				setTimeout(() => {
+					vscode.commands.executeCommand('cline-activator.activate');
+				}, 5000);
+			}
+			
+			function deactivate() {}
+			
+			module.exports = {
+				activate,
+				deactivate
+			}
+		`
+		fs.writeFileSync(extensionJsPath, extensionJs)
+		console.log(`Created Cline Activator extension`)
+
+		// Try multiple approaches to activate the extension
+		let serverStarted = false
+
+		// Create an activation script to run in VS Code
+		const activationScriptPath = path.join(settingsDir, "activate-cline.js")
+		const activationScript = `
+			// This script will be executed to activate Cline and start the test server
+			const vscode = require('vscode');
+			
+			// Execute the cline-activator.activate command
+			vscode.commands.executeCommand('cline-activator.activate');
+		`
+		fs.writeFileSync(activationScriptPath, activationScript)
+		console.log(`Created activation script to run in VS Code`)
+
+		// Execute the activation script
+		try {
+			console.log("Executing activation script to start Cline and test server...")
+			await execa(
+				"code",
+				[
+					"--user-data-dir",
+					tempUserDataDir,
+					"--extensions-dir",
+					tempExtensionsDir,
+					"--folder-uri",
+					`file://${workspacePath}`,
+					"--execute",
+					activationScriptPath,
+				],
+				{
+					stdio: "inherit",
+				},
+			)
+
+			// Wait for the test server to start
+			console.log("Waiting for test server to start...")
+			for (let i = 0; i < 30; i++) {
+				try {
+					// Try to connect to the test server
+					const response = await fetch("http://localhost:9876/task", {
+						method: "OPTIONS",
+						headers: {
+							"Content-Type": "application/json",
+						},
+					})
+
+					if (response.status === 204) {
+						console.log("Test server is running!")
+						serverStarted = true
+						break
+					}
+				} catch (error) {
+					// Server not started yet, wait and try again
+					await new Promise((resolve) => setTimeout(resolve, 1000))
+				}
+			}
+		} catch (error) {
+			console.warn("Failed to execute activation script:", error)
+		}
+
+		if (!serverStarted) {
+			console.warn("Test server did not start after multiple attempts")
+			console.log("You may need to manually open the Cline extension in VS Code")
+		}
+
+		// Store the resources for this workspace
+		const resources: VSCodeResources = {
+			tempUserDataDir,
+			tempExtensionsDir,
+		}
+
+		// Store in the global map
+		workspaceResources.set(workspacePath, resources)
+
+		// Return the resources
+		return resources
+	} catch (error: any) {
+		throw new Error(`Failed to spawn VSCode: ${error.message}`)
+	}
+}
+
+/**
+ * Clean up VS Code resources and shut down the test server
+ * @param workspacePath The workspace path to clean up resources for
+ */
+export async function cleanupVSCode(workspacePath: string): Promise<void> {
+	console.log(`Cleaning up VS Code resources for workspace: ${workspacePath}`)
+
+	// Get the resources for this workspace
+	const resources = workspaceResources.get(workspacePath)
+	if (!resources) {
+		console.log(`No resources found for workspace: ${workspacePath}`)
+		return
+	}
+
+	// Try to shut down the test server
+	try {
+		console.log("Shutting down test server...")
+		await fetch("http://localhost:9876/shutdown", {
+			method: "POST",
+			headers: {
+				"Content-Type": "application/json",
+			},
+		}).catch(() => {
+			// Ignore errors, the server might already be down
+		})
+	} catch (error) {
+		console.warn(`Error shutting down test server: ${error}`)
+	}
+
+	// Try to gracefully close VS Code instead of killing it
+	try {
+		console.log("Attempting to gracefully close VS Code...")
+
+		// Create a settings file that will disable the crash reporter and the exit confirmation dialog
+		const settingsDir = path.join(resources.tempUserDataDir, "User")
+		const settingsPath = path.join(settingsDir, "settings.json")
+
+		// Read existing settings if they exist
+		let settings = {}
+		if (fs.existsSync(settingsPath)) {
+			try {
+				settings = JSON.parse(fs.readFileSync(settingsPath, "utf8"))
+			} catch (error) {
+				console.warn(`Error reading settings file: ${error}`)
+			}
+		}
+
+		// Update settings to disable crash reporter and exit confirmation
+		settings = {
+			...settings,
+			"window.confirmBeforeClose": "never",
+			"telemetry.enableCrashReporter": false,
+			"window.restoreWindows": "none",
+			"window.newWindowDimensions": "default",
+		}
+
+		// Write updated settings
+		fs.writeFileSync(settingsPath, JSON.stringify(settings, null, 2))
+
+		// On macOS, use AppleScript to quit VS Code gracefully
+		if (process.platform === "darwin") {
+			try {
+				// First try AppleScript to quit VS Code gracefully
+				await execa("osascript", ["-e", 'tell application "Visual Studio Code" to quit'])
+
+				// Wait a moment for VS Code to close
+				await new Promise((resolve) => setTimeout(resolve, 2000))
+			} catch (appleScriptError) {
+				console.warn(`Error using AppleScript to quit VS Code: ${appleScriptError}`)
+			}
+		} else if (process.platform === "win32") {
+			// On Windows, try to use taskkill without /F first
+			try {
+				await execa("taskkill", ["/IM", "code.exe"])
+
+				// Wait a moment for VS Code to close
+				await new Promise((resolve) => setTimeout(resolve, 2000))
+			} catch (taskkillError) {
+				console.warn(`Error using taskkill to quit VS Code: ${taskkillError}`)
+			}
+		} else {
+			// On Linux, try to use SIGTERM first
+			try {
+				// Find VS Code processes
+				const { stdout } = await execa("ps", ["aux"])
+				const lines = stdout.split("\n")
+
+				for (const line of lines) {
+					if (line.includes(resources.tempUserDataDir)) {
+						const parts = line.trim().split(/\s+/)
+						const pid = parseInt(parts[1])
+
+						if (pid && !isNaN(pid)) {
+							console.log(`Sending SIGTERM to VS Code process with PID: ${pid}`)
+							try {
+								// Use SIGTERM instead of SIGKILL for a graceful shutdown
+								process.kill(pid, "SIGTERM")
+							} catch (killError) {
+								console.warn(`Failed to terminate process ${pid}: ${killError}`)
+							}
+						}
+					}
+				}
+
+				// Wait a moment for VS Code to close
+				await new Promise((resolve) => setTimeout(resolve, 2000))
+			} catch (psError) {
+				console.warn(`Error listing processes: ${psError}`)
+			}
+		}
+
+		// If graceful methods failed, fall back to forceful termination as a last resort
+		// Check if VS Code is still running with the temp user data dir
+		let vsCodeStillRunning = false
+
+		if (process.platform !== "win32") {
+			try {
+				const { stdout } = await execa("ps", ["aux"])
+				vsCodeStillRunning = stdout.split("\n").some((line) => line.includes(resources.tempUserDataDir))
+			} catch (error) {
+				console.warn(`Error checking if VS Code is still running: ${error}`)
+			}
+		} else {
+			try {
+				const { stdout } = await execa("tasklist", ["/FI", `IMAGENAME eq code.exe`])
+				vsCodeStillRunning = stdout.includes("code.exe")
+			} catch (error) {
+				console.warn(`Error checking if VS Code is still running: ${error}`)
+			}
+		}
+
+		// If VS Code is still running, use forceful termination as a last resort
+		if (vsCodeStillRunning) {
+			console.log("Graceful shutdown failed, falling back to forceful termination...")
+
+			if (process.platform === "win32") {
+				try {
+					await execa("taskkill", ["/IM", "code.exe", "/F"])
+				} catch (error) {
+					console.warn(`Error forcefully terminating VS Code: ${error}`)
+				}
+			} else {
+				try {
+					const { stdout } = await execa("ps", ["aux"])
+					const lines = stdout.split("\n")
+
+					for (const line of lines) {
+						if (line.includes(resources.tempUserDataDir)) {
+							const parts = line.trim().split(/\s+/)
+							const pid = parseInt(parts[1])
+
+							if (pid && !isNaN(pid)) {
+								console.log(`Forcefully killing VS Code process with PID: ${pid}`)
+								try {
+									process.kill(pid, "SIGKILL")
+								} catch (killError) {
+									console.warn(`Failed to kill process ${pid}: ${killError}`)
+								}
+							}
+						}
+					}
+				} catch (error) {
+					console.warn(`Error forcefully terminating VS Code: ${error}`)
+				}
+			}
+		}
+	} catch (error) {
+		console.warn(`Error closing VS Code: ${error}`)
+	}
+
+	// Clean up temporary directories
+	try {
+		console.log(`Removing temporary user data directory: ${resources.tempUserDataDir}`)
+		fs.rmSync(resources.tempUserDataDir, { recursive: true, force: true })
+	} catch (error) {
+		console.warn(`Error removing temporary user data directory: ${error}`)
+	}
+
+	try {
+		console.log(`Removing temporary extensions directory: ${resources.tempExtensionsDir}`)
+		fs.rmSync(resources.tempExtensionsDir, { recursive: true, force: true })
+	} catch (error) {
+		console.warn(`Error removing temporary extensions directory: ${error}`)
+	}
+
+	// Remove from the global map
+	workspaceResources.delete(workspacePath)
+
+	console.log("Cleanup completed")
+}

evals/cli/tsconfig.json

@@ -0,0 +1,17 @@
+{
+	"compilerOptions": {
+		"target": "ES2020",
+		"module": "commonjs",
+		"lib": ["ES2020"],
+		"declaration": true,
+		"outDir": "dist",
+		"rootDir": "src",
+		"strict": true,
+		"esModuleInterop": true,
+		"skipLibCheck": true,
+		"forceConsistentCasingInFileNames": true,
+		"resolveJsonModule": true
+	},
+	"include": ["src/**/*"],
+	"exclude": ["node_modules", "dist"]
+}

locales/ar-sa/README.md

@@ -2,7 +2,7 @@
 العربية | <a href="https://github.com/cline/cline/blob/main/locales/es/README.md" target="_blank">الإسبانية</a> | <a href="https://github.com/cline/cline/blob/main/locales/de/README.md" target="_blank">الألمانية</a> | <a href="https://github.com/cline/cline/blob/main/locales/ja/README.md" target="_blank">اليابانية</a> | <a href="https://github.com/cline/cline/blob/main/locales/zh-cn/README.md" target="_blank">الصينية المبسطة</a> | <a href="https://github.com/cline/cline/blob/main/locales/zh-tw/README.md" target="_blank">الصينية التقليدية</a> | <a href="https://github.com/cline/cline/blob/main/locales/pt-BR/README.md" target="_blank">البرتغالية</a>
 </sub></div>
 
-# Cline – \#1 على OpenRouter
+# Cline
 
 <p align="center">
   <img src="https://media.githubusercontent.com/media/cline/cline/main/assets/docs/demo.gif" width="100%" />

locales/de/README.md

@@ -1,4 +1,4 @@
-# Cline – \#1 auf OpenRouter
+# Cline
 
 <p align="center">
         <img src="https://media.githubusercontent.com/media/cline/cline/main/assets/docs/demo.gif" width="100%" />

locales/es/README.md

@@ -1,4 +1,4 @@
-# Cline – #1 en OpenRouter
+# Cline
 
 <p align="center">
         <img src="https://media.githubusercontent.com/media/cline/cline/main/assets/docs/demo.gif" width="100%" />

locales/ja/CONTRIBUTING.md

@@ -1,4 +1,4 @@
-# Clineへの貢献
+# Cline
 
 Clineへの貢献に興味をお持ちいただきありがとうございます。
 

locales/ja/README.md

@@ -1,4 +1,4 @@
-# Cline – OpenRouterでのナンバーワン
+# Cline
 
 <p align="center">
     <img src="https://media.githubusercontent.com/media/cline/cline/main/assets/docs/demo.gif" width="100%" />

locales/ko/CONTRIBUTING.md

@@ -1,4 +1,4 @@
-# Cline에 기여하기
+# Cline
 
 Cline에 기여하는 것에 관심을 가져주셔서 감사합니다! 버그 수정, 기능 추가, 문서 개선 등 모든 기여는 Cline을 더욱 스마트하게 만드는 데 기여합니다. 활기차고 환영하는 커뮤니티를 유지하기 위해 모든 구성원은 [행동 강령](CODE_OF_CONDUCT.md)을 준수해야 합니다.
 

locales/ko/README.md

@@ -1,4 +1,4 @@
-# Cline - 최고의 OpenRouter
+# Cline
 
 <p align="center">
     <img src="https://media.githubusercontent.com/media/cline/cline/main/assets/docs/demo.gif" width="100%" />

locales/pt-BR/README.md

@@ -1,4 +1,4 @@
-# Cline – #1 no OpenRouter
+# Cline
 
 <p align="center">
         <img src="https://media.githubusercontent.com/media/cline/cline/main/assets/docs/demo.gif" width="100%" />

locales/zh-cn/README.md

@@ -1,4 +1,4 @@
-# Cline – OpenRouter 排名第一的 AI 工具
+# Cline
 
 <p align="center">
     <img src="https://media.githubusercontent.com/media/cline/cline/main/assets/docs/demo.gif" width="100%" />

locales/zh-tw/README.md

@@ -2,7 +2,7 @@
 <a href="https://github.com/cline/cline/blob/main/README.md" target="_blank">English</a> | <a href="https://github.com/cline/cline/blob/main/locales/es/README.md" target="_blank">Español</a> | <a href="https://github.com/cline/cline/blob/main/locales/de/README.md" target="_blank">Deutsch</a> | <a href="https://github.com/cline/cline/blob/main/locales/ja/README.md" target="_blank">日本語</a> | <a href="https://github.com/cline/cline/blob/main/locales/zh-cn/README.md" target="_blank">简体中文</a> | 繁體中文 | <a href="https://github.com/cline/cline/blob/main/locales/ko/README.md" target="_blank">한국어</a>
 </sub></div>
 
-# Cline – OpenRouter 第一名的 AI 工具
+# Cline
 
 <p align="center">
     <img src="https://media.githubusercontent.com/media/cline/cline/main/assets/docs/demo.gif" width="100%" />

package-lock.json

@@ -1,20 +1,22 @@
 {
 	"name": "claude-dev",
-	"version": "3.10.1",
+	"version": "3.12.3",
 	"lockfileVersion": 3,
 	"requires": true,
 	"packages": {
 		"": {
 			"name": "claude-dev",
-			"version": "3.10.1",
+			"version": "3.12.3",
 			"license": "Apache-2.0",
 			"dependencies": {
 				"@anthropic-ai/bedrock-sdk": "^0.12.4",
 				"@anthropic-ai/sdk": "^0.37.0",
 				"@anthropic-ai/vertex-sdk": "^0.6.4",
 				"@aws-sdk/client-bedrock-runtime": "^3.758.0",
+				"@bufbuild/protobuf": "^2.2.5",
 				"@google-cloud/vertexai": "^1.9.3",
 				"@google/generative-ai": "^0.18.0",
+				"@grpc/grpc-js": "^1.9.15",
 				"@mistralai/mistralai": "^1.5.0",
 				"@modelcontextprotocol/sdk": "^1.7.0",
 				"@opentelemetry/api": "^1.4.1",
@@ -79,15 +81,18 @@
 				"@vscode/test-cli": "^0.0.9",
 				"@vscode/test-electron": "^2.4.0",
 				"chai": "^4.3.10",
+				"chalk": "^5.3.0",
 				"esbuild": "^0.25.0",
 				"eslint": "^8.57.0",
 				"husky": "^9.1.7",
 				"npm-run-all": "^4.1.5",
 				"prettier": "^3.3.3",
+				"protoc-gen-ts": "^0.8.7",
 				"proxyquire": "^2.1.3",
 				"should": "^13.2.3",
 				"sinon": "^19.0.2",
 				"ts-node": "^10.9.2",
+				"ts-proto": "^2.6.1",
 				"typescript": "^5.4.5"
 			},
 			"engines": {
@@ -3957,6 +3962,12 @@
 			"dev": true,
 			"license": "MIT"
 		},
+		"node_modules/@bufbuild/protobuf": {
+			"version": "2.2.5",
+			"resolved": "https://registry.npmjs.org/@bufbuild/protobuf/-/protobuf-2.2.5.tgz",
+			"integrity": "sha512-/g5EzJifw5GF8aren8wZ/G5oMuPoGeS6MQD3ca8ddcvdXR5UELUfdTZITCGNhNXynY/AYl3Z4plmxdj/tRl/hQ==",
+			"license": "(Apache-2.0 AND BSD-3-Clause)"
+		},
 		"node_modules/@changesets/apply-release-plan": {
 			"version": "7.0.8",
 			"resolved": "https://registry.npmjs.org/@changesets/apply-release-plan/-/apply-release-plan-7.0.8.tgz",
@@ -5566,6 +5577,7 @@
 			"version": "1.9.15",
 			"resolved": "https://registry.npmjs.org/@grpc/grpc-js/-/grpc-js-1.9.15.tgz",
 			"integrity": "sha512-nqE7Hc0AzI+euzUwDAy0aY5hCp10r734gMGRdU+qOPX0XSceI2ULrcXB5U2xSc5VkWwalCj4M7GzCAygZl2KoQ==",
+			"license": "Apache-2.0",
 			"dependencies": {
 				"@grpc/proto-loader": "^0.7.8",
 				"@types/node": ">=12.12.47"
@@ -9919,6 +9931,19 @@
 				"url": "https://github.com/sponsors/sindresorhus"
 			}
 		},
+		"node_modules/case-anything": {
+			"version": "2.1.13",
+			"resolved": "https://registry.npmjs.org/case-anything/-/case-anything-2.1.13.tgz",
+			"integrity": "sha512-zlOQ80VrQ2Ue+ymH5OuM/DlDq64mEm+B9UTdHULv5osUMD6HalNTblf2b1u/m6QecjsnOkBpqVZ+XPwIVsy7Ng==",
+			"dev": true,
+			"license": "MIT",
+			"engines": {
+				"node": ">=12.13"
+			},
+			"funding": {
+				"url": "https://github.com/sponsors/mesqueeb"
+			}
+		},
 		"node_modules/chai": {
 			"version": "4.3.10",
 			"resolved": "https://registry.npmjs.org/chai/-/chai-4.3.10.tgz",
@@ -9938,35 +9963,18 @@
 			}
 		},
 		"node_modules/chalk": {
-			"version": "4.1.2",
-			"resolved": "https://registry.npmjs.org/chalk/-/chalk-4.1.2.tgz",
-			"integrity": "sha512-oKnbhFyRIXpUuez8iBMmyEa4nbj4IOQyuhc/wy9kY7/WVPcwIO9VA668Pu8RkO7+0G76SLROeyw9CpQ061i4mA==",
+			"version": "5.4.1",
+			"resolved": "https://registry.npmjs.org/chalk/-/chalk-5.4.1.tgz",
+			"integrity": "sha512-zgVZuo2WcZgfUEmsn6eO3kINexW8RAE4maiQ8QNs8CtpPCSyMiYsULR3HQYkm3w8FIA3SberyMJMSldGsW+U3w==",
 			"dev": true,
 			"license": "MIT",
-			"dependencies": {
-				"ansi-styles": "^4.1.0",
-				"supports-color": "^7.1.0"
-			},
 			"engines": {
-				"node": ">=10"
+				"node": "^12.17.0 || ^14.13 || >=16.0.0"
 			},
 			"funding": {
 				"url": "https://github.com/chalk/chalk?sponsor=1"
 			}
 		},
-		"node_modules/chalk/node_modules/supports-color": {
-			"version": "7.2.0",
-			"resolved": "https://registry.npmjs.org/supports-color/-/supports-color-7.2.0.tgz",
-			"integrity": "sha512-qpCAvRl9stuOHveKsn7HncJRvv501qIacKzQlO/+Lwxc9+0q2wLyv4Dfvt80/DPn2pqOBsJdDiogXGR9+OvwRw==",
-			"dev": true,
-			"license": "MIT",
-			"dependencies": {
-				"has-flag": "^4.0.0"
-			},
-			"engines": {
-				"node": ">=8"
-			}
-		},
 		"node_modules/chardet": {
 			"version": "0.7.0",
 			"resolved": "https://registry.npmjs.org/chardet/-/chardet-0.7.0.tgz",
@@ -10612,6 +10620,19 @@
 				"node": ">=8"
 			}
 		},
+		"node_modules/detect-libc": {
+			"version": "1.0.3",
+			"resolved": "https://registry.npmjs.org/detect-libc/-/detect-libc-1.0.3.tgz",
+			"integrity": "sha512-pGjwhsmsp4kL2RTz08wcOlGN83otlqHeD/Z5T8GXZB+/YcpQ/dgo+lbU8ZsGxV0HIvqqxo9l7mqYwyYMD9bKDg==",
+			"dev": true,
+			"license": "Apache-2.0",
+			"bin": {
+				"detect-libc": "bin/detect-libc.js"
+			},
+			"engines": {
+				"node": ">=0.10"
+			}
+		},
 		"node_modules/devtools-protocol": {
 			"version": "0.0.1342118",
 			"resolved": "https://registry.npmjs.org/devtools-protocol/-/devtools-protocol-0.0.1342118.tgz",
@@ -10722,6 +10743,16 @@
 				"url": "https://github.com/fb55/domutils?sponsor=1"
 			}
 		},
+		"node_modules/dprint-node": {
+			"version": "1.0.8",
+			"resolved": "https://registry.npmjs.org/dprint-node/-/dprint-node-1.0.8.tgz",
+			"integrity": "sha512-iVKnUtYfGrYcW1ZAlfR/F59cUVL8QIhWoBJoSjkkdua/dkWIgjZfiLMeTjiB06X0ZLkQ0M2C1VbUj/CxkIf1zg==",
+			"dev": true,
+			"license": "MIT",
+			"dependencies": {
+				"detect-libc": "^1.0.3"
+			}
+		},
 		"node_modules/duck": {
 			"version": "0.1.12",
 			"resolved": "https://registry.npmjs.org/duck/-/duck-0.1.12.tgz",
@@ -11192,6 +11223,23 @@
 				"concat-map": "0.0.1"
 			}
 		},
+		"node_modules/eslint/node_modules/chalk": {
+			"version": "4.1.2",
+			"resolved": "https://registry.npmjs.org/chalk/-/chalk-4.1.2.tgz",
+			"integrity": "sha512-oKnbhFyRIXpUuez8iBMmyEa4nbj4IOQyuhc/wy9kY7/WVPcwIO9VA668Pu8RkO7+0G76SLROeyw9CpQ061i4mA==",
+			"dev": true,
+			"license": "MIT",
+			"dependencies": {
+				"ansi-styles": "^4.1.0",
+				"supports-color": "^7.1.0"
+			},
+			"engines": {
+				"node": ">=10"
+			},
+			"funding": {
+				"url": "https://github.com/chalk/chalk?sponsor=1"
+			}
+		},
 		"node_modules/eslint/node_modules/glob-parent": {
 			"version": "6.0.2",
 			"resolved": "https://registry.npmjs.org/glob-parent/-/glob-parent-6.0.2.tgz",
@@ -11240,6 +11288,19 @@
 				"node": ">=8"
 			}
 		},
+		"node_modules/eslint/node_modules/supports-color": {
+			"version": "7.2.0",
+			"resolved": "https://registry.npmjs.org/supports-color/-/supports-color-7.2.0.tgz",
+			"integrity": "sha512-qpCAvRl9stuOHveKsn7HncJRvv501qIacKzQlO/+Lwxc9+0q2wLyv4Dfvt80/DPn2pqOBsJdDiogXGR9+OvwRw==",
+			"dev": true,
+			"license": "MIT",
+			"dependencies": {
+				"has-flag": "^4.0.0"
+			},
+			"engines": {
+				"node": ">=8"
+			}
+		},
 		"node_modules/espree": {
 			"version": "9.6.1",
 			"resolved": "https://registry.npmjs.org/espree/-/espree-9.6.1.tgz",
@@ -13659,6 +13720,36 @@
 				"url": "https://github.com/sponsors/sindresorhus"
 			}
 		},
+		"node_modules/log-symbols/node_modules/chalk": {
+			"version": "4.1.2",
+			"resolved": "https://registry.npmjs.org/chalk/-/chalk-4.1.2.tgz",
+			"integrity": "sha512-oKnbhFyRIXpUuez8iBMmyEa4nbj4IOQyuhc/wy9kY7/WVPcwIO9VA668Pu8RkO7+0G76SLROeyw9CpQ061i4mA==",
+			"dev": true,
+			"license": "MIT",
+			"dependencies": {
+				"ansi-styles": "^4.1.0",
+				"supports-color": "^7.1.0"
+			},
+			"engines": {
+				"node": ">=10"
+			},
+			"funding": {
+				"url": "https://github.com/chalk/chalk?sponsor=1"
+			}
+		},
+		"node_modules/log-symbols/node_modules/supports-color": {
+			"version": "7.2.0",
+			"resolved": "https://registry.npmjs.org/supports-color/-/supports-color-7.2.0.tgz",
+			"integrity": "sha512-qpCAvRl9stuOHveKsn7HncJRvv501qIacKzQlO/+Lwxc9+0q2wLyv4Dfvt80/DPn2pqOBsJdDiogXGR9+OvwRw==",
+			"dev": true,
+			"license": "MIT",
+			"dependencies": {
+				"has-flag": "^4.0.0"
+			},
+			"engines": {
+				"node": ">=8"
+			}
+		},
 		"node_modules/long": {
 			"version": "5.2.4",
 			"resolved": "https://registry.npmjs.org/long/-/long-5.2.4.tgz",
@@ -14768,19 +14859,6 @@
 				"url": "https://github.com/sponsors/sindresorhus"
 			}
 		},
-		"node_modules/ora/node_modules/chalk": {
-			"version": "5.3.0",
-			"resolved": "https://registry.npmjs.org/chalk/-/chalk-5.3.0.tgz",
-			"integrity": "sha512-dLitG79d+GV1Nb/VYcCDFivJeK1hiukt9QjRNVOsUtTy1rR1YJsmpGGTZ3qJos+uw7WmWF4wUwBd9jxjocFC2w==",
-			"dev": true,
-			"license": "MIT",
-			"engines": {
-				"node": "^12.17.0 || ^14.13 || >=16.0.0"
-			},
-			"funding": {
-				"url": "https://github.com/chalk/chalk?sponsor=1"
-			}
-		},
 		"node_modules/ora/node_modules/emoji-regex": {
 			"version": "10.3.0",
 			"resolved": "https://registry.npmjs.org/emoji-regex/-/emoji-regex-10.3.0.tgz",
@@ -15366,6 +15444,20 @@
 				"node": ">=12.0.0"
 			}
 		},
+		"node_modules/protoc-gen-ts": {
+			"version": "0.8.7",
+			"resolved": "https://registry.npmjs.org/protoc-gen-ts/-/protoc-gen-ts-0.8.7.tgz",
+			"integrity": "sha512-jr4VJey2J9LVYCV7EVyVe53g1VMw28cCmYJhBe5e3YX5wiyiDwgxWxeDf9oTqAe4P1bN/YGAkW2jhlH8LohwiQ==",
+			"dev": true,
+			"license": "MIT",
+			"bin": {
+				"protoc-gen-ts": "protoc-gen-ts.js"
+			},
+			"funding": {
+				"type": "individual",
+				"url": "https://www.buymeacoffee.com/thesayyn"
+			}
+		},
 		"node_modules/proxy-addr": {
 			"version": "2.0.7",
 			"resolved": "https://registry.npmjs.org/proxy-addr/-/proxy-addr-2.0.7.tgz",
@@ -16970,6 +17062,42 @@
 				"node": ">=0.3.1"
 			}
 		},
+		"node_modules/ts-poet": {
+			"version": "6.11.0",
+			"resolved": "https://registry.npmjs.org/ts-poet/-/ts-poet-6.11.0.tgz",
+			"integrity": "sha512-r5AGF8vvb+GjBsnqiTqbLhN1/U2FJt6BI+k0dfCrkKzWvUhNlwMmq9nDHuucHs45LomgHjZPvYj96dD3JawjJA==",
+			"dev": true,
+			"license": "Apache-2.0",
+			"dependencies": {
+				"dprint-node": "^1.0.8"
+			}
+		},
+		"node_modules/ts-proto": {
+			"version": "2.7.0",
+			"resolved": "https://registry.npmjs.org/ts-proto/-/ts-proto-2.7.0.tgz",
+			"integrity": "sha512-BGHjse2wTOeswOqnnPKinpxmbaRd882so/e1En6ww59YMG7AO9Kg4vPpJcbVfrpBixPRDqHafXD/RDyd2T99GA==",
+			"dev": true,
+			"license": "ISC",
+			"dependencies": {
+				"@bufbuild/protobuf": "^2.0.0",
+				"case-anything": "^2.1.13",
+				"ts-poet": "^6.7.0",
+				"ts-proto-descriptors": "2.0.0"
+			},
+			"bin": {
+				"protoc-gen-ts_proto": "protoc-gen-ts_proto"
+			}
+		},
+		"node_modules/ts-proto-descriptors": {
+			"version": "2.0.0",
+			"resolved": "https://registry.npmjs.org/ts-proto-descriptors/-/ts-proto-descriptors-2.0.0.tgz",
+			"integrity": "sha512-wHcTH3xIv11jxgkX5OyCSFfw27agpInAd6yh89hKG6zqIXnjW9SYqSER2CVQxdPj4czeOhGagNvZBEbJPy7qkw==",
+			"dev": true,
+			"license": "ISC",
+			"dependencies": {
+				"@bufbuild/protobuf": "^2.0.0"
+			}
+		},
 		"node_modules/tslib": {
 			"version": "1.14.1",
 			"resolved": "https://registry.npmjs.org/tslib/-/tslib-1.14.1.tgz",

package.json

@@ -2,7 +2,7 @@
 	"name": "claude-dev",
 	"displayName": "Cline",
 	"description": "Autonomous coding agent right in your IDE, capable of creating/editing files, running commands, using the browser, and more with your permission every step of the way.",
-	"version": "3.11.1",
+	"version": "3.12.3",
 	"icon": "assets/icons/icon.png",
 	"engines": {
 		"vscode": "^1.84.0"
@@ -116,6 +116,11 @@
 				"command": "cline.fixWithCline",
 				"title": "Fix with Cline",
 				"category": "Cline"
+			},
+			{
+				"command": "cline.focusChatInput",
+				"title": "Jump to Chat Input",
+				"category": "Cline"
 			}
 		],
 		"keybindings": [
@@ -224,21 +229,6 @@
 					},
 					"description": "Settings for VSCode Language Model API"
 				},
-				"cline.mcp.mode": {
-					"type": "string",
-					"enum": [
-						"full",
-						"server-use-only",
-						"off"
-					],
-					"enumDescriptions": [
-						"Enable all MCP functionality (server use and build instructions)",
-						"Enable MCP server use only (excludes instructions about building MCP servers)",
-						"Disable all MCP functionality"
-					],
-					"default": "full",
-					"description": "Controls MCP inclusion in prompts, reduces token usage if you only need access to certain functionality."
-				},
 				"cline.enableCheckpoints": {
 					"type": "boolean",
 					"default": true,
@@ -305,6 +295,7 @@
 		"watch:esbuild:test": "IS_TEST=true node esbuild.js --watch",
 		"watch:tsc": "tsc --noEmit --watch --project tsconfig.json",
 		"package": "npm run build:webview && npm run check-types && npm run lint && node esbuild.js --production",
+		"protos": "node proto/build-proto.js && prettier src/shared/proto --write && prettier src/core/controller --write",
 		"package:test": "IS_TEST=true npm run build:webview:test && npm run check-types && npm run lint && IS_TEST=true node esbuild.js --production",
 		"build:webview:test": "cd webview-ui && IS_TEST=true npm run build",
 		"watch:test": "IS_TEST=true npm-run-all -p watch:tsc watch:esbuild:test",
@@ -349,15 +340,18 @@
 		"@vscode/test-cli": "^0.0.9",
 		"@vscode/test-electron": "^2.4.0",
 		"chai": "^4.3.10",
+		"chalk": "^5.3.0",
 		"esbuild": "^0.25.0",
 		"eslint": "^8.57.0",
 		"husky": "^9.1.7",
 		"npm-run-all": "^4.1.5",
 		"prettier": "^3.3.3",
+		"protoc-gen-ts": "^0.8.7",
 		"proxyquire": "^2.1.3",
 		"should": "^13.2.3",
 		"sinon": "^19.0.2",
 		"ts-node": "^10.9.2",
+		"ts-proto": "^2.6.1",
 		"typescript": "^5.4.5"
 	},
 	"dependencies": {
@@ -365,8 +359,10 @@
 		"@anthropic-ai/sdk": "^0.37.0",
 		"@anthropic-ai/vertex-sdk": "^0.6.4",
 		"@aws-sdk/client-bedrock-runtime": "^3.758.0",
+		"@bufbuild/protobuf": "^2.2.5",
 		"@google-cloud/vertexai": "^1.9.3",
 		"@google/generative-ai": "^0.18.0",
+		"@grpc/grpc-js": "^1.9.15",
 		"@mistralai/mistralai": "^1.5.0",
 		"@modelcontextprotocol/sdk": "^1.7.0",
 		"@opentelemetry/api": "^1.4.1",

proto/browser.proto

@@ -0,0 +1,22 @@
+syntax = "proto3";
+
+package cline;
+
+import "common.proto";
+
+service BrowserService {
+  rpc getBrowserConnectionInfo(EmptyRequest) returns (BrowserConnectionInfo);
+  rpc testBrowserConnection(StringRequest) returns (BrowserConnection);
+}
+
+message BrowserConnectionInfo {
+  bool is_connected = 1;
+  bool is_remote = 2;
+  optional string host = 3;
+}
+
+message BrowserConnection {
+  bool success = 1;
+  string message = 2;
+  optional string endpoint = 3;
+}

proto/build-proto.js

@@ -0,0 +1,157 @@
+#!/usr/bin/env node
+
+import * as fs from "fs/promises"
+import * as path from "path"
+import { execSync } from "child_process"
+import { globby } from "globby"
+import chalk from "chalk"
+
+// Get script directory and root directory
+const SCRIPT_DIR = path.dirname(new URL(import.meta.url).pathname)
+const ROOT_DIR = path.resolve(SCRIPT_DIR, "..")
+
+async function main() {
+	console.log(chalk.bold.blue("Starting Protocol Buffer code generation..."))
+
+	// Check if protoc is installed
+	try {
+		const options = { stdio: "ignore" }
+		execSync("protoc --version", options)
+	} catch (error) {
+		console.warn(chalk.yellow("Warning: protoc is not installed. Skipping proto generation."))
+		console.warn(chalk.yellow("To install Protocol Buffers compiler, visit: https://grpc.io/docs/protoc-installation/"))
+		process.exit(0) // Exit with success as requested
+	}
+
+	// Check if ts-proto plugin is available
+	const TS_PROTO_PLUGIN = path.join(ROOT_DIR, "node_modules", ".bin", "protoc-gen-ts_proto")
+	try {
+		await fs.access(TS_PROTO_PLUGIN)
+	} catch (error) {
+		console.error(chalk.red("Error: ts-proto plugin not found at"), TS_PROTO_PLUGIN)
+		console.error(chalk.red('Please run "npm install" to install the required dependencies.'))
+		process.exit(1)
+	}
+
+	// Define output directories
+	const TS_OUT_DIR = path.join(ROOT_DIR, "src", "shared", "proto")
+
+	// Create output directory if it doesn't exist
+	await fs.mkdir(TS_OUT_DIR, { recursive: true })
+
+	// Clean up existing generated files
+	console.log(chalk.cyan("Cleaning up existing generated TypeScript files..."))
+	const existingFiles = await globby("**/*.ts", { cwd: TS_OUT_DIR })
+	for (const file of existingFiles) {
+		await fs.unlink(path.join(TS_OUT_DIR, file))
+	}
+
+	// Process all proto files
+	console.log(chalk.cyan("Processing proto files from"), SCRIPT_DIR)
+	const protoFiles = await globby("**/*.proto", { cwd: SCRIPT_DIR })
+
+	for (const protoFile of protoFiles) {
+		console.log(chalk.cyan(`Generating TypeScript code for ${protoFile}...`))
+
+		// Build the protoc command with proper path handling for cross-platform
+		const protocCommand = [
+			"protoc",
+			`--plugin=protoc-gen-ts_proto="${TS_PROTO_PLUGIN}"`,
+			`--ts_proto_out="${TS_OUT_DIR}"`,
+			"--ts_proto_opt=outputServices=generic-definitions,env=node,esModuleInterop=true,useDate=false,useOptionals=messages",
+			`--proto_path="${SCRIPT_DIR}"`,
+			`"${path.join(SCRIPT_DIR, protoFile)}"`,
+		].join(" ")
+
+		try {
+			const execOptions = {
+				stdio: "inherit",
+			}
+			execSync(protocCommand, execOptions)
+		} catch (error) {
+			console.error(chalk.red(`Error generating TypeScript for ${protoFile}:`), error)
+			process.exit(1)
+		}
+	}
+
+	console.log(chalk.green("Protocol Buffer code generation completed successfully."))
+	console.log(chalk.green(`TypeScript files generated in: ${TS_OUT_DIR}`))
+
+	// Generate method registration files
+	await generateMethodRegistrations()
+
+	// Make the script executable
+	try {
+		await fs.chmod(path.join(SCRIPT_DIR, "build-proto.js"), 0o755)
+	} catch (error) {
+		console.warn(chalk.yellow("Warning: Could not make script executable:"), error)
+	}
+}
+
+async function generateMethodRegistrations() {
+	console.log(chalk.cyan("Generating method registration files..."))
+
+	const serviceDirs = [
+		path.join(ROOT_DIR, "src", "core", "controller", "browser"),
+		// Add more service directories here as needed
+	]
+
+	for (const serviceDir of serviceDirs) {
+		try {
+			await fs.access(serviceDir)
+		} catch (error) {
+			console.log(chalk.gray(`Skipping ${serviceDir} - directory does not exist`))
+			continue
+		}
+
+		const serviceName = path.basename(serviceDir)
+		const registryFile = path.join(serviceDir, "methods.ts")
+
+		console.log(chalk.cyan(`Generating method registrations for ${serviceName}...`))
+
+		// Get all TypeScript files in the service directory
+		const files = await globby("*.ts", { cwd: serviceDir })
+
+		// Filter out index.ts and methods.ts
+		const implementationFiles = files.filter((file) => file !== "index.ts" && file !== "methods.ts")
+
+		// Create the output file with header
+		let content = `// AUTO-GENERATED FILE - DO NOT MODIFY DIRECTLY
+// Generated by proto/build-proto.js
+
+// Import all method implementations
+import { registerMethod } from "./index"\n`
+
+		// Add imports for all implementation files
+		for (const file of implementationFiles) {
+			const baseName = path.basename(file, ".ts")
+			content += `import { ${baseName} } from "./${baseName}"\n`
+		}
+
+		// Add registration function
+		content += `\n// Register all ${serviceName} service methods
+export function registerAllMethods(): void {
+\t// Register each method with the registry\n`
+
+		// Add registration statements
+		for (const file of implementationFiles) {
+			const baseName = path.basename(file, ".ts")
+			content += `\tregisterMethod("${baseName}", ${baseName})\n`
+		}
+
+		// Close the function
+		content += `}`
+
+		// Write the file
+		await fs.writeFile(registryFile, content)
+		console.log(chalk.green(`Generated ${registryFile}`))
+	}
+
+	console.log(chalk.green("Method registration files generated successfully."))
+}
+
+// Run the main function
+main().catch((error) => {
+	console.error(chalk.red("Error:"), error)
+	process.exit(1)
+})

proto/common.proto

@@ -0,0 +1,40 @@
+syntax = "proto3";
+
+package cline;
+
+message Metadata {
+}
+
+message EmptyRequest {
+  Metadata metadata = 1;
+}
+
+message Empty {
+}
+
+message StringRequest {
+  Metadata metadata = 1;
+  string value = 2;
+}
+
+message String {
+  string value = 1;
+}
+
+message Int64Request {
+  Metadata metadata = 1;
+  int64 value = 2;
+}
+
+message Int64 {
+  int64 value = 1;
+}
+
+message BytesRequest {
+  Metadata metadata = 1;
+  bytes value = 2;
+}
+
+message Bytes {
+  bytes value = 1;
+}

proto/package.json

@@ -0,0 +1,3 @@
+{
+	"type": "module"
+}

src/api/providers/openai-native.ts

@@ -12,7 +12,7 @@ import {
 import { convertToOpenAiMessages } from "../transform/openai-format"
 import { calculateApiCostOpenAI } from "../../utils/cost"
 import { ApiStream } from "../transform/stream"
-import { ChatCompletionReasoningEffort } from "openai/resources/chat/completions.mjs"
+import type { ChatCompletionReasoningEffort } from "openai/resources/chat/completions"
 
 export class OpenAiNativeHandler implements ApiHandler {
 	private options: ApiHandlerOptions
@@ -50,7 +50,7 @@ export class OpenAiNativeHandler implements ApiHandler {
 			case "o1":
 			case "o1-preview":
 			case "o1-mini": {
-				// o1 doesnt support streaming, non-1 temp, or system prompt
+				// o1 doesn't support streaming, non-1 temp, or system prompt
 				const response = await this.client.chat.completions.create({
 					model: model.id,
 					messages: [{ role: "user", content: systemPrompt }, ...convertToOpenAiMessages(messages)],
@@ -64,6 +64,8 @@ export class OpenAiNativeHandler implements ApiHandler {
 
 				break
 			}
+			case "o4-mini":
+			case "o3":
 			case "o3-mini": {
 				const stream = await this.client.chat.completions.create({
 					model: model.id,

src/api/providers/openai.ts

@@ -6,7 +6,7 @@ import { ApiHandler } from "../index"
 import { convertToOpenAiMessages } from "../transform/openai-format"
 import { ApiStream } from "../transform/stream"
 import { convertToR1Format } from "../transform/r1-format"
-import { ChatCompletionReasoningEffort } from "openai/resources/chat/completions.mjs"
+import type { ChatCompletionReasoningEffort } from "openai/resources/chat/completions"
 
 export class OpenAiHandler implements ApiHandler {
 	private options: ApiHandlerOptions
@@ -35,7 +35,7 @@ export class OpenAiHandler implements ApiHandler {
 		const modelId = this.options.openAiModelId ?? ""
 		const isDeepseekReasoner = modelId.includes("deepseek-reasoner")
 		const isR1FormatRequired = this.options.openAiModelInfo?.isR1FormatRequired ?? false
-		const isO3Mini = modelId.includes("o3-mini")
+		const isReasoningModelFamily = modelId.includes("o3") || modelId.includes("o4")
 
 		let openAiMessages: OpenAI.Chat.ChatCompletionMessageParam[] = [
 			{ role: "system", content: systemPrompt },
@@ -55,7 +55,7 @@ export class OpenAiHandler implements ApiHandler {
 			openAiMessages = convertToR1Format([{ role: "user", content: systemPrompt }, ...messages])
 		}
 
-		if (isO3Mini) {
+		if (isReasoningModelFamily) {
 			openAiMessages = [{ role: "developer", content: systemPrompt }, ...convertToOpenAiMessages(messages)]
 			temperature = undefined // does not support temperature
 			reasoningEffort = (this.options.o3MiniReasoningEffort as ChatCompletionReasoningEffort) || "medium"

src/api/providers/types.ts

@@ -1,6 +1,13 @@
 // For the following openrouter error type sources, see the docs here:
 // https://openrouter.ai/docs/api-reference/errors
 
+export interface LanguageModelChatSelector {
+	vendor?: string
+	family?: string
+	version?: string
+	id?: string
+}
+
 export type OpenRouterErrorResponse = {
 	error: {
 		message: string

src/api/providers/vertex.ts

@@ -5,6 +5,7 @@ import { ApiHandler } from "../"
 import { ApiHandlerOptions, ModelInfo, vertexDefaultModelId, VertexModelId, vertexModels } from "../../shared/api"
 import { ApiStream } from "../transform/stream"
 import { VertexAI } from "@google-cloud/vertexai"
+import { calculateApiCostOpenAI } from "../../utils/cost"
 
 // https://docs.anthropic.com/en/api/claude-on-vertex-ai
 export class VertexHandler implements ApiHandler {
@@ -266,6 +267,17 @@ export class VertexHandler implements ApiHandler {
 					}
 				}
 			}
+			// Handle token usage metadata
+			const { usageMetadata } = await streamingResult.response
+			if (usageMetadata) {
+				const { promptTokenCount = 0, candidatesTokenCount = 0 } = usageMetadata
+				yield {
+					type: "usage",
+					inputTokens: promptTokenCount,
+					outputTokens: candidatesTokenCount,
+					totalCost: calculateApiCostOpenAI(model.info, promptTokenCount, candidatesTokenCount, 0, 0),
+				}
+			}
 		}
 	}
 

src/api/providers/vscode-lm.ts

@@ -6,6 +6,7 @@ import { ApiStream } from "../transform/stream"
 import { convertToVsCodeLmMessages } from "../transform/vscode-lm-format"
 import { SELECTOR_SEPARATOR, stringifyVsCodeLmModelSelector } from "../../shared/vsCodeSelectorUtils"
 import { ApiHandlerOptions, ModelInfo, openAiModelInfoSaneDefaults } from "../../shared/api"
+import type { LanguageModelChatSelector as LanguageModelChatSelectorFromTypes } from "./types"
 
 // Cline does not update VSCode type definitions or engine requirements to maintain compatibility.
 // This declaration (as seen in src/integrations/TerminalManager.ts) provides types for the Language Model API in newer versions of VSCode.
@@ -19,12 +20,7 @@ declare module "vscode" {
 		Auto = 1,
 		Required = 2,
 	}
-	interface LanguageModelChatSelector {
-		vendor?: string
-		family?: string
-		version?: string
-		id?: string
-	}
+	interface LanguageModelChatSelector extends LanguageModelChatSelectorFromTypes {}
 	interface LanguageModelChatTool {
 		name: string
 		description: string

src/api/providers/xai.ts

@@ -4,7 +4,7 @@ import { ApiHandler } from "../"
 import { ApiHandlerOptions, XAIModelId, ModelInfo, xaiDefaultModelId, xaiModels } from "../../shared/api"
 import { convertToOpenAiMessages } from "../transform/openai-format"
 import { ApiStream } from "../transform/stream"
-import { ChatCompletionReasoningEffort } from "openai/resources/chat/completions.mjs"
+import { ChatCompletionReasoningEffort } from "openai/resources/chat/completions"
 
 export class XAIHandler implements ApiHandler {
 	private options: ApiHandlerOptions
@@ -34,6 +34,7 @@ export class XAIHandler implements ApiHandler {
 			temperature: 0,
 			messages: [{ role: "system", content: systemPrompt }, ...convertToOpenAiMessages(messages)],
 			stream: true,
+			stream_options: { include_usage: true },
 			reasoning_effort: reasoningEffort,
 		})
 
@@ -46,7 +47,7 @@ export class XAIHandler implements ApiHandler {
 				}
 			}
 
-			if ("reasoning_content" in delta && delta.reasoning_content) {
+			if (delta && "reasoning_content" in delta && delta.reasoning_content) {
 				yield {
 					type: "reasoning",
 					// @ts-ignore-next-line

src/api/transform/vscode-lm-format.test.ts

@@ -0,0 +1,211 @@
+// This file contains `declare module "vscode"` so we must import it.
+import "../providers/vscode-lm"
+import { describe, it } from "mocha"
+import "should"
+import * as vscode from "vscode"
+import { Anthropic } from "@anthropic-ai/sdk"
+import { asObjectSafe, convertToAnthropicRole, convertToVsCodeLmMessages, convertToAnthropicMessage } from "./vscode-lm-format"
+
+describe("asObjectSafe", () => {
+	it("should handle falsy values", () => {
+		asObjectSafe(0).should.deepEqual({})
+		asObjectSafe("").should.deepEqual({})
+		asObjectSafe(null).should.deepEqual({})
+		asObjectSafe(undefined).should.deepEqual({})
+	})
+
+	it("should parse valid JSON strings", () => {
+		asObjectSafe('{"key": "value"}').should.deepEqual({ key: "value" })
+	})
+
+	it("should return an empty object for invalid JSON strings", () => {
+		asObjectSafe("invalid json").should.deepEqual({})
+	})
+
+	it("should convert objects to plain objects", () => {
+		const input = { prop: "value" }
+		asObjectSafe(input).should.deepEqual(input)
+		asObjectSafe(input).should.not.equal(input) // Should be a new object
+	})
+
+	it("should convert arrays to plain objects", () => {
+		const input = ["hello world"]
+		asObjectSafe(input).should.deepEqual({ 0: "hello world" })
+	})
+})
+
+describe("convertToAnthropicRole", () => {
+	it("should convert VSCode roles to Anthropic roles", () => {
+		// @ts-expect-error（Testing with an invalid role）
+		const unknownRole = "unknown" as vscode.LanguageModelChatMessageRole
+		;(convertToAnthropicRole(vscode.LanguageModelChatMessageRole.Assistant) === "assistant").should.be.true()
+		;(convertToAnthropicRole(vscode.LanguageModelChatMessageRole.User) === "user").should.be.true()
+		;(convertToAnthropicRole(unknownRole) === null).should.be.true()
+	})
+})
+
+describe("convertToVsCodeLmMessages", () => {
+	it("should convert simple string messages", () => {
+		const anthropicMessages: Anthropic.Messages.MessageParam[] = [
+			{ role: "user", content: "Hello" },
+			{ role: "assistant", content: "Hi there" },
+		]
+
+		const result = convertToVsCodeLmMessages(anthropicMessages)
+
+		result.should.have.length(2)
+		result[0].role.should.equal(vscode.LanguageModelChatMessageRole.User)
+		result[0].content[0].should.be.instanceof(vscode.LanguageModelTextPart)
+		const textPart0 = result[0].content[0] as vscode.LanguageModelTextPart
+		textPart0.should.have.property("value", "Hello")
+
+		result[1].role.should.equal(vscode.LanguageModelChatMessageRole.Assistant)
+		result[1].content[0].should.be.instanceof(vscode.LanguageModelTextPart)
+		const textPart1 = result[1].content[0] as vscode.LanguageModelTextPart
+		textPart1.should.have.property("value", "Hi there")
+	})
+
+	it("should convert complex user messages with tool results", () => {
+		const anthropicMessages: Anthropic.Messages.MessageParam[] = [
+			{
+				role: "user",
+				content: [
+					{ type: "text", text: "User text" },
+					{
+						type: "tool_result",
+						tool_use_id: "tool-123",
+						content: [{ type: "text", text: "Tool result" }],
+					},
+				],
+			},
+		]
+
+		const result = convertToVsCodeLmMessages(anthropicMessages)
+
+		result.should.have.length(1)
+		result[0].role.should.equal(vscode.LanguageModelChatMessageRole.User)
+		result[0].content.should.have.length(2)
+
+		// Check that the first content part is a ToolResultPart
+		result[0].content[0].should.be.instanceof(vscode.LanguageModelToolResultPart)
+		const toolResultPart = result[0].content[0] as vscode.LanguageModelToolResultPart
+		toolResultPart.should.have.property("callId", "tool-123")
+
+		// Skip detailed testing of internal structure as it may vary
+		// Just verify it's the right type with the right ID
+
+		// Check the second content part is a TextPart
+		result[0].content[1].should.be.instanceof(vscode.LanguageModelTextPart)
+		const textPart = result[0].content[1] as vscode.LanguageModelTextPart
+		textPart.should.have.property("value", "User text")
+	})
+
+	it("should convert complex assistant messages with tool calls", () => {
+		const anthropicMessages: Anthropic.Messages.MessageParam[] = [
+			{
+				role: "assistant",
+				content: [
+					{ type: "text", text: "Assistant text" },
+					{
+						type: "tool_use",
+						id: "tool-123",
+						name: "testTool",
+						input: { param: "value" },
+					},
+				],
+			},
+		]
+
+		const result = convertToVsCodeLmMessages(anthropicMessages)
+
+		result.should.have.length(1)
+		result[0].role.should.equal(vscode.LanguageModelChatMessageRole.Assistant)
+		result[0].content.should.have.length(2)
+
+		result[0].content[0].should.be.instanceof(vscode.LanguageModelToolCallPart)
+		const toolCallPart = result[0].content[0] as vscode.LanguageModelToolCallPart
+		toolCallPart.should.have.property("callId", "tool-123")
+		toolCallPart.should.have.property("name", "testTool")
+		toolCallPart.should.have.property("input")
+		toolCallPart.input.should.deepEqual({ param: "value" })
+
+		result[0].content[1].should.be.instanceof(vscode.LanguageModelTextPart)
+		const textPart = result[0].content[1] as vscode.LanguageModelTextPart
+		textPart.should.have.property("value", "Assistant text")
+	})
+
+	it("should handle image blocks with appropriate placeholders", () => {
+		const anthropicMessages: Anthropic.Messages.MessageParam[] = [
+			{
+				role: "user",
+				content: [
+					{
+						type: "image",
+						source: {
+							type: "base64",
+							media_type: "image/jpeg",
+							data: "base64data",
+						},
+					},
+				],
+			},
+		]
+
+		const result = convertToVsCodeLmMessages(anthropicMessages)
+
+		result.should.have.length(1)
+		result[0].content[0].should.be.instanceof(vscode.LanguageModelTextPart)
+		const textPart = result[0].content[0] as vscode.LanguageModelTextPart
+		textPart.should.have.property("value")
+		textPart.value.should.match(/Image \(base64\): image\/jpeg not supported by VSCode LM API/)
+	})
+})
+
+describe("convertToAnthropicMessage", () => {
+	it("should convert VSCode assistant messages to Anthropic format", () => {
+		const vsCodeMsg = vscode.LanguageModelChatMessage.Assistant([
+			new vscode.LanguageModelTextPart("Test message"),
+			new vscode.LanguageModelToolCallPart("tool-id", "testTool", { param: "value" }),
+		])
+
+		const result = convertToAnthropicMessage(vsCodeMsg)
+
+		result.should.have.property("role", "assistant")
+		result.should.have.property("content").which.is.an.Array()
+		result.content.should.have.length(2)
+
+		// Check properties carefully to avoid null reference errors
+		if (result.content && result.content.length >= 1) {
+			const textContent = result.content[0]
+			if (textContent) {
+				textContent.should.have.property("type", "text")
+				if (textContent.type === "text") {
+					textContent.should.have.property("text", "Test message")
+				}
+			}
+		}
+
+		if (result.content && result.content.length >= 2) {
+			const toolContent = result.content[1]
+			if (toolContent) {
+				toolContent.should.have.property("type", "tool_use")
+				if (toolContent.type === "tool_use") {
+					toolContent.should.have.property("id", "tool-id")
+					toolContent.should.have.property("name", "testTool")
+					toolContent.should.have.property("input").which.deepEqual({ param: "value" })
+				}
+			}
+		}
+	})
+
+	it("should throw an error for non-assistant messages", () => {
+		const vsCodeMsg = vscode.LanguageModelChatMessage.User("User message")
+
+		try {
+			convertToAnthropicMessage(vsCodeMsg)
+			throw new Error("Should have thrown an error")
+		} catch (error: any) {
+			error.message.should.match(/Only assistant messages are supported/)
+		}
+	})
+})

src/api/transform/vscode-lm-format.ts

@@ -4,7 +4,7 @@ import * as vscode from "vscode"
 /**
  * Safely converts a value into a plain object.
  */
-function asObjectSafe(value: any): object {
+export function asObjectSafe(value: any): object {
 	// Handle null/undefined
 	if (!value) {
 		return {}
@@ -145,7 +145,9 @@ export function convertToVsCodeLmMessages(
 	return vsCodeLmMessages
 }
 
-export function convertToAnthropicRole(vsCodeLmMessageRole: vscode.LanguageModelChatMessageRole): string | null {
+export function convertToAnthropicRole(
+	vsCodeLmMessageRole: vscode.LanguageModelChatMessageRole,
+): Anthropic.Messages.MessageParam["role"] | null {
 	switch (vsCodeLmMessageRole) {
 		case vscode.LanguageModelChatMessageRole.Assistant:
 			return "assistant"
@@ -156,10 +158,8 @@ export function convertToAnthropicRole(vsCodeLmMessageRole: vscode.LanguageModel
 	}
 }
 
-export async function convertToAnthropicMessage(
-	vsCodeLmMessage: vscode.LanguageModelChatMessage,
-): Promise<Anthropic.Messages.Message> {
-	const anthropicRole: string | null = convertToAnthropicRole(vsCodeLmMessage.role)
+export function convertToAnthropicMessage(vsCodeLmMessage: vscode.LanguageModelChatMessage): Anthropic.Messages.Message {
+	const anthropicRole = convertToAnthropicRole(vsCodeLmMessage.role)
 	if (anthropicRole !== "assistant") {
 		throw new Error("Cline <Language Model API>: Only assistant messages are supported.")
 	}

src/core/assistant-message/index.ts

@@ -21,6 +21,7 @@ export const toolUseNames = [
 	"access_mcp_resource",
 	"ask_followup_question",
 	"plan_mode_respond",
+	"load_mcp_documentation",
 	"attempt_completion",
 	"new_task",
 ] as const

src/core/assistant-message/parse-assistant-message.ts

@@ -123,7 +123,7 @@ export function parseAssistantMessage(assistantMessage: string) {
 		contentBlocks.push(currentToolUse)
 	}
 
-	// Note: it doesnt matter if check for currentToolUse or currentTextContent, only one of them will be defined since only one can be partial at a time
+	// Note: it doesn't matter if check for currentToolUse or currentTextContent, only one of them will be defined since only one can be partial at a time
 	if (currentTextContent) {
 		// stream did not complete text content, add it as partial
 		contentBlocks.push(currentTextContent)

src/core/context/context-management/ContextManager-legacy.ts

@@ -0,0 +1,102 @@
+import { Anthropic } from "@anthropic-ai/sdk"
+import { ClineApiReqInfo, ClineMessage } from "../../../shared/ExtensionMessage"
+import { ApiHandler } from "../../../api"
+import { getContextWindowInfo } from "./context-window-utils"
+
+class ContextManager {
+	getNewContextMessagesAndMetadata(
+		apiConversationHistory: Anthropic.Messages.MessageParam[],
+		clineMessages: ClineMessage[],
+		api: ApiHandler,
+		conversationHistoryDeletedRange: [number, number] | undefined,
+		previousApiReqIndex: number,
+	) {
+		let updatedConversationHistoryDeletedRange = false
+
+		// If the previous API request's total token usage is close to the context window, truncate the conversation history to free up space for the new request
+		if (previousApiReqIndex >= 0) {
+			const previousRequest = clineMessages[previousApiReqIndex]
+			if (previousRequest && previousRequest.text) {
+				const { tokensIn, tokensOut, cacheWrites, cacheReads }: ClineApiReqInfo = JSON.parse(previousRequest.text)
+				const totalTokens = (tokensIn || 0) + (tokensOut || 0) + (cacheWrites || 0) + (cacheReads || 0)
+				const { maxAllowedSize } = getContextWindowInfo(api)
+
+				// This is the most reliable way to know when we're close to hitting the context window.
+				if (totalTokens >= maxAllowedSize) {
+					// Since the user may switch between models with different context windows, truncating half may not be enough (ie if switching from claude 200k to deepseek 64k, half truncation will only remove 100k tokens, but we need to remove much more)
+					// So if totalTokens/2 is greater than maxAllowedSize, we truncate 3/4 instead of 1/2
+					// FIXME: truncating the conversation in a way that is optimal for prompt caching AND takes into account multi-context window complexity is something we need to improve
+					const keep = totalTokens / 2 > maxAllowedSize ? "quarter" : "half"
+
+					// NOTE: it's okay that we overwriteConversationHistory in resume task since we're only ever removing the last user message and not anything in the middle which would affect this range
+					conversationHistoryDeletedRange = this.getNextTruncationRange(
+						apiConversationHistory,
+						conversationHistoryDeletedRange,
+						keep,
+					)
+
+					updatedConversationHistoryDeletedRange = true
+				}
+			}
+		}
+
+		// conversationHistoryDeletedRange is updated only when we're close to hitting the context window, so we don't continuously break the prompt cache
+		const truncatedConversationHistory = this.getTruncatedMessages(apiConversationHistory, conversationHistoryDeletedRange)
+
+		return {
+			conversationHistoryDeletedRange: conversationHistoryDeletedRange,
+			updatedConversationHistoryDeletedRange: updatedConversationHistoryDeletedRange,
+			truncatedConversationHistory: truncatedConversationHistory,
+		}
+	}
+
+	public getNextTruncationRange(
+		apiMessages: Anthropic.Messages.MessageParam[],
+		currentDeletedRange: [number, number] | undefined,
+		keep: "half" | "quarter",
+	): [number, number] {
+		// Since we always keep the first message, currentDeletedRange[0] will always be 1 (for now until we have a smarter truncation algorithm)
+		const rangeStartIndex = 1
+		const startOfRest = currentDeletedRange ? currentDeletedRange[1] + 1 : 1
+
+		let messagesToRemove: number
+		if (keep === "half") {
+			// Remove half of remaining user-assistant pairs
+			// We first calculate half of the messages then divide by 2 to get the number of pairs.
+			// After flooring, we multiply by 2 to get the number of messages.
+			// Note that this will also always be an even number.
+			messagesToRemove = Math.floor((apiMessages.length - startOfRest) / 4) * 2 // Keep even number
+		} else {
+			// Remove 3/4 of remaining user-assistant pairs
+			// We calculate 3/4ths of the messages then divide by 2 to get the number of pairs.
+			// After flooring, we multiply by 2 to get the number of messages.
+			// Note that this will also always be an even number.
+			messagesToRemove = Math.floor(((apiMessages.length - startOfRest) * 3) / 4 / 2) * 2
+		}
+
+		let rangeEndIndex = startOfRest + messagesToRemove - 1
+
+		// Make sure the last message being removed is a user message, so that the next message after the initial task message is an assistant message. This preservers the user-assistant-user-assistant structure.
+		// NOTE: anthropic format messages are always user-assistant-user-assistant, while openai format messages can have multiple user messages in a row (we use anthropic format throughout cline)
+		if (apiMessages[rangeEndIndex].role !== "user") {
+			rangeEndIndex -= 1
+		}
+
+		// this is an inclusive range that will be removed from the conversation history
+		return [rangeStartIndex, rangeEndIndex]
+	}
+
+	public getTruncatedMessages(
+		messages: Anthropic.Messages.MessageParam[],
+		deletedRange: [number, number] | undefined,
+	): Anthropic.Messages.MessageParam[] {
+		if (!deletedRange) {
+			return messages
+		}
+
+		const [start, end] = deletedRange
+		// the range is inclusive - both start and end indices and everything in between will be removed from the final result.
+		// NOTE: if you try to console log these, don't forget that logging a reference to an array may not provide the same result as logging a slice() snapshot of that array at that exact moment. The following DOES in fact include the latest assistant message.
+		return [...messages.slice(0, start), ...messages.slice(end + 1)]
+	}
+}

src/core/context/context-management/ContextManager-new.ts

@@ -1,834 +0,0 @@
-import { getContextWindowInfo } from "./context-window-utils"
-import { formatResponse } from "../../prompts/responses"
-import { GlobalFileNames } from "../../storage/disk"
-import { fileExistsAtPath } from "../../../utils/fs"
-import * as path from "path"
-import fs from "fs/promises"
-import cloneDeep from "clone-deep"
-import { ClineApiReqInfo, ClineMessage } from "../../../shared/ExtensionMessage"
-import { ApiHandler } from "../../../api"
-import { Anthropic } from "@anthropic-ai/sdk"
-
-enum EditType {
-	UNDEFINED = 0,
-	NO_FILE_READ = 1,
-	READ_FILE_TOOL = 2,
-	ALTER_FILE_TOOL = 3,
-	FILE_MENTION = 4,
-}
-
-// array of string values allows us to cover all changes for message types currently supported
-type MessageContent = string[]
-type MessageMetadata = string[][]
-
-// Type for a single context update
-type ContextUpdate = [number, string, MessageContent, MessageMetadata] // [timestamp, updateType, update, metadata]
-
-// Type for the serialized format of our nested maps
-type SerializedContextHistory = Array<
-	[
-		number, // messageIndex
-		[
-			number, // EditType (message type)
-			Array<
-				[
-					number, // blockIndex
-					ContextUpdate[], // updates array (now with 4 elements including metadata)
-				]
-			>,
-		],
-	]
->
-
-export class ContextManager {
-	// mapping from the apiMessages outer index to the inner message index to a list of actual changes, ordered by timestamp
-	// timestamp is required in order to support full checkpointing, where the changes we apply need to be able to be undone when
-	// moving to an earlier conversation history checkpoint - this ordering intuitively allows for binary search on truncation
-	// there is also a number stored for each (EditType) which defines which message type it is, for custom handling
-
-	// format:  { outerIndex => [EditType, { innerIndex => [[timestamp, updateType, update], ...] }] }
-	// example: { 1 => { [0, 0 => [[<timestamp>, "text", "[NOTE] Some previous conversation history with the user has been removed ..."], ...] }] }
-	// the above example would be how we update the first assistant message to indicate we truncated text
-	private contextHistoryUpdates: Map<number, [number, Map<number, ContextUpdate[]>]>
-
-	constructor() {
-		this.contextHistoryUpdates = new Map()
-	}
-
-	/**
-	 * public function for loading contextHistoryUpdates from disk, if it exists
-	 */
-	async initializeContextHistory(taskDirectory: string) {
-		this.contextHistoryUpdates = await this.getSavedContextHistory(taskDirectory)
-	}
-
-	/**
-	 * get the stored context history updates from disk
-	 */
-	private async getSavedContextHistory(taskDirectory: string): Promise<Map<number, [number, Map<number, ContextUpdate[]>]>> {
-		try {
-			const filePath = path.join(taskDirectory, GlobalFileNames.contextHistory)
-			if (await fileExistsAtPath(filePath)) {
-				const data = await fs.readFile(filePath, "utf8")
-				const serializedUpdates = JSON.parse(data) as SerializedContextHistory
-
-				// Update to properly reconstruct the tuple structure
-				return new Map(
-					serializedUpdates.map(([messageIndex, [numberValue, innerMapArray]]) => [
-						messageIndex,
-						[numberValue, new Map(innerMapArray)],
-					]),
-				)
-			}
-		} catch (error) {
-			console.error("Failed to load context history:", error)
-		}
-		return new Map()
-	}
-
-	/**
-	 * save the context history updates to disk
-	 */
-	private async saveContextHistory(taskDirectory: string) {
-		try {
-			const serializedUpdates: SerializedContextHistory = Array.from(this.contextHistoryUpdates.entries()).map(
-				([messageIndex, [numberValue, innerMap]]) => [messageIndex, [numberValue, Array.from(innerMap.entries())]],
-			)
-
-			await fs.writeFile(
-				path.join(taskDirectory, GlobalFileNames.contextHistory),
-				JSON.stringify(serializedUpdates),
-				"utf8",
-			)
-		} catch (error) {
-			console.error("Failed to save context history:", error)
-		}
-	}
-
-	/**
-	 * primary entry point for getting up to date context & truncating when required
-	 */
-	async getNewContextMessagesAndMetadata(
-		apiConversationHistory: Anthropic.Messages.MessageParam[],
-		clineMessages: ClineMessage[],
-		api: ApiHandler,
-		conversationHistoryDeletedRange: [number, number] | undefined,
-		previousApiReqIndex: number,
-		taskDirectory: string,
-	) {
-		let updatedConversationHistoryDeletedRange = false
-
-		// If the previous API request's total token usage is close to the context window, truncate the conversation history to free up space for the new request
-		if (previousApiReqIndex >= 0) {
-			const previousRequest = clineMessages[previousApiReqIndex]
-			if (previousRequest && previousRequest.text) {
-				const timestamp = previousRequest.ts
-				const { tokensIn, tokensOut, cacheWrites, cacheReads }: ClineApiReqInfo = JSON.parse(previousRequest.text)
-				const totalTokens = (tokensIn || 0) + (tokensOut || 0) + (cacheWrites || 0) + (cacheReads || 0)
-				const { maxAllowedSize } = getContextWindowInfo(api)
-
-				// This is the most reliable way to know when we're close to hitting the context window.
-				if (totalTokens >= maxAllowedSize) {
-					// Since the user may switch between models with different context windows, truncating half may not be enough (ie if switching from claude 200k to deepseek 64k, half truncation will only remove 100k tokens, but we need to remove much more)
-					// So if totalTokens/2 is greater than maxAllowedSize, we truncate 3/4 instead of 1/2
-					const keep = totalTokens / 2 > maxAllowedSize ? "quarter" : "half"
-
-					// we later check how many chars we trim to determine if we should still truncate history
-					let [anyContextUpdates, uniqueFileReadIndices] = this.applyContextOptimizations(
-						apiConversationHistory,
-						conversationHistoryDeletedRange ? conversationHistoryDeletedRange[1] + 1 : 2,
-						timestamp,
-					)
-
-					let needToTruncate = true
-					if (anyContextUpdates) {
-						// determine whether we've saved enough chars to not truncate
-						const charactersSavedPercentage = this.calculateContextOptimizationMetrics(
-							apiConversationHistory,
-							conversationHistoryDeletedRange,
-							uniqueFileReadIndices,
-						)
-						if (charactersSavedPercentage >= 0.3) {
-							needToTruncate = false
-						}
-					}
-
-					if (needToTruncate) {
-						// go ahead with truncation
-						anyContextUpdates = this.applyStandardContextTruncationNoticeChange(timestamp) || anyContextUpdates
-
-						// NOTE: it's okay that we overwriteConversationHistory in resume task since we're only ever removing the last user message and not anything in the middle which would affect this range
-						conversationHistoryDeletedRange = this.getNextTruncationRange(
-							apiConversationHistory,
-							conversationHistoryDeletedRange,
-							keep,
-						)
-
-						updatedConversationHistoryDeletedRange = true
-					}
-
-					// if we alter the context history, save the updated version to disk
-					if (anyContextUpdates) {
-						await this.saveContextHistory(taskDirectory)
-					}
-				}
-			}
-		}
-
-		const truncatedConversationHistory = this.getAndAlterTruncatedMessages(
-			apiConversationHistory,
-			conversationHistoryDeletedRange,
-		)
-
-		return {
-			conversationHistoryDeletedRange: conversationHistoryDeletedRange,
-			updatedConversationHistoryDeletedRange: updatedConversationHistoryDeletedRange,
-			truncatedConversationHistory: truncatedConversationHistory,
-		}
-	}
-
-	/**
-	 * get truncation range
-	 */
-	public getNextTruncationRange(
-		apiMessages: Anthropic.Messages.MessageParam[],
-		currentDeletedRange: [number, number] | undefined,
-		keep: "half" | "quarter",
-	): [number, number] {
-		// We always keep the first user-assistant pairing, and truncate an even number of messages from there
-		const rangeStartIndex = 2 // index 0 and 1 are kept
-		const startOfRest = currentDeletedRange ? currentDeletedRange[1] + 1 : 2 // inclusive starting index
-
-		let messagesToRemove: number
-		if (keep === "half") {
-			// Remove half of remaining user-assistant pairs
-			// We first calculate half of the messages then divide by 2 to get the number of pairs.
-			// After flooring, we multiply by 2 to get the number of messages.
-			// Note that this will also always be an even number.
-			messagesToRemove = Math.floor((apiMessages.length - startOfRest) / 4) * 2 // Keep even number
-		} else {
-			// Remove 3/4 of remaining user-assistant pairs
-			// We calculate 3/4ths of the messages then divide by 2 to get the number of pairs.
-			// After flooring, we multiply by 2 to get the number of messages.
-			// Note that this will also always be an even number.
-			messagesToRemove = Math.floor(((apiMessages.length - startOfRest) * 3) / 4 / 2) * 2
-		}
-
-		let rangeEndIndex = startOfRest + messagesToRemove - 1 // inclusive ending index
-
-		// Make sure that the last message being removed is a assistant message, so the next message after the initial user-assistant pair is an assistant message. This preserves the user-assistant-user-assistant structure.
-		// NOTE: anthropic format messages are always user-assistant-user-assistant, while openai format messages can have multiple user messages in a row (we use anthropic format throughout cline)
-		if (apiMessages[rangeEndIndex].role !== "assistant") {
-			rangeEndIndex -= 1
-		}
-
-		// this is an inclusive range that will be removed from the conversation history
-		return [rangeStartIndex, rangeEndIndex]
-	}
-
-	/**
-	 * external interface to support old calls
-	 */
-	public getTruncatedMessages(
-		messages: Anthropic.Messages.MessageParam[],
-		deletedRange: [number, number] | undefined,
-	): Anthropic.Messages.MessageParam[] {
-		return this.getAndAlterTruncatedMessages(messages, deletedRange)
-	}
-
-	/**
-	 * apply all required truncation methods to the messages in context
-	 */
-	private getAndAlterTruncatedMessages(
-		messages: Anthropic.Messages.MessageParam[],
-		deletedRange: [number, number] | undefined,
-	): Anthropic.Messages.MessageParam[] {
-		if (messages.length <= 1) {
-			return messages
-		}
-
-		const updatedMessages = this.applyContextHistoryUpdates(messages, deletedRange ? deletedRange[1] + 1 : 2)
-
-		// OLD NOTE: if you try to console log these, don't forget that logging a reference to an array may not provide the same result as logging a slice() snapshot of that array at that exact moment. The following DOES in fact include the latest assistant message.
-		return updatedMessages
-	}
-
-	/**
-	 * applies deletedRange truncation and other alterations based on changes in this.contextHistoryUpdates
-	 */
-	private applyContextHistoryUpdates(
-		messages: Anthropic.Messages.MessageParam[],
-		startFromIndex: number,
-	): Anthropic.Messages.MessageParam[] {
-		// runtime is linear in length of user messages, if expecting a limited number of alterations, could be more optimal to loop over alterations
-
-		const firstChunk = messages.slice(0, 2) // get first user-assistant pair
-		const secondChunk = messages.slice(startFromIndex) // get remaining messages within context
-		const messagesToUpdate = [...firstChunk, ...secondChunk]
-
-		// we need the mapping from the local indices in messagesToUpdate to the global array of updates in this.contextHistoryUpdates
-		const originalIndices = [
-			...Array(2).keys(),
-			...Array(secondChunk.length)
-				.fill(0)
-				.map((_, i) => i + startFromIndex),
-		]
-
-		for (let arrayIndex = 0; arrayIndex < messagesToUpdate.length; arrayIndex++) {
-			const messageIndex = originalIndices[arrayIndex]
-
-			const innerTuple = this.contextHistoryUpdates.get(messageIndex)
-			if (!innerTuple) {
-				continue
-			}
-
-			// because we are altering this, we need a deep copy
-			messagesToUpdate[arrayIndex] = cloneDeep(messagesToUpdate[arrayIndex])
-
-			// Extract the map from the tuple
-			const innerMap = innerTuple[1]
-			for (const [blockIndex, changes] of innerMap) {
-				// apply the latest change among n changes - [timestamp, updateType, update]
-				const latestChange = changes[changes.length - 1]
-
-				if (latestChange[1] === "text") {
-					// only altering text for now
-					const message = messagesToUpdate[arrayIndex]
-
-					if (Array.isArray(message.content)) {
-						const block = message.content[blockIndex]
-						if (block && block.type === "text") {
-							block.text = latestChange[2][0]
-						}
-					}
-				}
-			}
-		}
-
-		return messagesToUpdate
-	}
-
-	/**
-	 * removes all context history updates that occurred after the specified timestamp and saves to disk
-	 */
-	async truncateContextHistory(timestamp: number, taskDirectory: string): Promise<void> {
-		this.truncateContextHistoryAtTimestamp(this.contextHistoryUpdates, timestamp)
-
-		// save the modified context history to disk
-		await this.saveContextHistory(taskDirectory)
-	}
-
-	/**
-	 * alters the context history to remove all alterations after a given timestamp
-	 * removes the index if there are no alterations there anymore, both outer and inner indices
-	 */
-	private truncateContextHistoryAtTimestamp(
-		contextHistory: Map<number, [number, Map<number, ContextUpdate[]>]>,
-		timestamp: number,
-	): void {
-		for (const [messageIndex, [_, innerMap]] of contextHistory) {
-			// track which blockIndices to delete
-			const blockIndicesToDelete: number[] = []
-
-			// loop over the innerIndices of the messages in this block
-			for (const [blockIndex, updates] of innerMap) {
-				// updates ordered by timestamp, so find cutoff point by iterating from right to left
-				let cutoffIndex = updates.length - 1
-				while (cutoffIndex >= 0 && updates[cutoffIndex][0] > timestamp) {
-					cutoffIndex--
-				}
-
-				// If we found updates to remove
-				if (cutoffIndex < updates.length - 1) {
-					// Modify the array in place to keep only updates up to cutoffIndex
-					updates.length = cutoffIndex + 1
-
-					// If no updates left after truncation, mark this block for deletion
-					if (updates.length === 0) {
-						blockIndicesToDelete.push(blockIndex)
-					}
-				}
-			}
-
-			// Remove empty blocks from inner map
-			for (const blockIndex of blockIndicesToDelete) {
-				innerMap.delete(blockIndex)
-			}
-
-			// If inner map is now empty, remove the message index from outer map
-			if (innerMap.size === 0) {
-				contextHistory.delete(messageIndex)
-			}
-		}
-	}
-
-	/**
-	 * applies the context optimization steps and returns whether any changes were made
-	 */
-	private applyContextOptimizations(
-		apiMessages: Anthropic.Messages.MessageParam[],
-		startFromIndex: number,
-		timestamp: number,
-	): [boolean, Set<number>] {
-		const [fileReadUpdatesBool, uniqueFileReadIndices] = this.findAndPotentiallySaveFileReadContextHistoryUpdates(
-			apiMessages,
-			startFromIndex,
-			timestamp,
-		)
-
-		// true if any context optimization steps alter state
-		const contextHistoryUpdated = fileReadUpdatesBool
-
-		return [contextHistoryUpdated, uniqueFileReadIndices]
-	}
-
-	/**
-	 * if there is any truncation and there is no other alteration already set, alter the assistant message to indicate this occurred
-	 */
-	private applyStandardContextTruncationNoticeChange(timestamp: number): boolean {
-		if (!this.contextHistoryUpdates.has(1)) {
-			// first assistant message always at index 1
-			const innerMap = new Map<number, ContextUpdate[]>()
-			innerMap.set(0, [[timestamp, "text", [formatResponse.contextTruncationNotice()], []]])
-			this.contextHistoryUpdates.set(1, [0, innerMap]) // EditType is undefined for first assistant message
-			return true
-		}
-		return false
-	}
-
-	/**
-	 * wraps the logic for determining file reads to overwrite, and altering state
-	 * returns whether any updates were made (bool) and indices where updates were made
-	 */
-	private findAndPotentiallySaveFileReadContextHistoryUpdates(
-		apiMessages: Anthropic.Messages.MessageParam[],
-		startFromIndex: number,
-		timestamp: number,
-	): [boolean, Set<number>] {
-		const [fileReadIndices, messageFilePaths] = this.getPossibleDuplicateFileReads(apiMessages, startFromIndex)
-		return this.applyFileReadContextHistoryUpdates(fileReadIndices, messageFilePaths, apiMessages, timestamp)
-	}
-
-	/**
-	 * generate a mapping from unique file reads from multiple tool calls to their outer index position(s)
-	 * also return additional metadata to support multiple file reads in file mention text blocks
-	 */
-	private getPossibleDuplicateFileReads(
-		apiMessages: Anthropic.Messages.MessageParam[],
-		startFromIndex: number,
-	): [Map<string, [number, number, string, string][]>, Map<number, string[]>] {
-		// fileReadIndices: { fileName => [outerIndex, EditType, searchText, replaceText] }
-		// messageFilePaths: { outerIndex => [fileRead1, fileRead2, ..] }
-		// searchText in fileReadIndices is only required for file mention file-reads since there can be more than one file in the text
-		// searchText will be the empty string "" in the case that it's not required, for non-file mentions
-		// messageFilePaths is only used for file mentions as there can be multiple files read in the same text chunk
-
-		// for all text blocks per file, has info for updating the block
-		const fileReadIndices = new Map<string, [number, number, string, string][]>()
-
-		// for file mention text blocks, track all the unique files read
-		const messageFilePaths = new Map<number, string[]>()
-
-		for (let i = startFromIndex; i < apiMessages.length; i++) {
-			let thisExistingFileReads: string[] = []
-
-			if (this.contextHistoryUpdates.has(i)) {
-				const innerTuple = this.contextHistoryUpdates.get(i)
-
-				if (innerTuple) {
-					// safety check
-					const editType = innerTuple[0]
-
-					if (editType === EditType.FILE_MENTION) {
-						const innerMap = innerTuple[1]
-
-						const blockIndex = 1 // file mention blocks assumed to be at index 1
-						const blockUpdates = innerMap.get(blockIndex)
-
-						// if we have updated this text previously, we want to check whether the lists of files in the metadata are the same
-						if (blockUpdates && blockUpdates.length > 0) {
-							// the first list indicates the files we have replaced in this text, second list indicates all unique files in this text
-							// if they are equal then we have replaced all the files in this text already, and can ignore further processing
-							if (
-								blockUpdates[blockUpdates.length - 1][3][0].length ===
-								blockUpdates[blockUpdates.length - 1][3][1].length
-							) {
-								continue
-							}
-							// otherwise there are still file reads here we can overwrite, so still need to process this text chunk
-							// to do so we need to keep track of which files we've already replaced so we don't replace them again
-							else {
-								thisExistingFileReads = blockUpdates[blockUpdates.length - 1][3][0]
-							}
-						}
-					} else {
-						// for all other cases we can assume that we dont need to check this again
-						continue
-					}
-				}
-			}
-
-			const message = apiMessages[i]
-			if (message.role === "user" && Array.isArray(message.content) && message.content.length > 0) {
-				const firstBlock = message.content[0]
-				if (firstBlock.type === "text") {
-					const matchTup = this.parsePotentialToolCall(firstBlock.text)
-					let foundNormalFileRead = false
-					if (matchTup) {
-						if (matchTup[0] === "read_file") {
-							this.handleReadFileToolCall(i, matchTup[1], fileReadIndices)
-							foundNormalFileRead = true
-						} else if (matchTup[0] === "replace_in_file" || matchTup[0] === "write_to_file") {
-							if (message.content.length > 1) {
-								const secondBlock = message.content[1]
-								if (secondBlock.type === "text") {
-									this.handlePotentialFileChangeToolCalls(i, matchTup[1], secondBlock.text, fileReadIndices)
-									foundNormalFileRead = true
-								}
-							}
-						}
-					}
-
-					// file mentions can happen in most other user message blocks
-					if (!foundNormalFileRead) {
-						if (message.content.length > 1) {
-							const secondBlock = message.content[1]
-							if (secondBlock.type === "text") {
-								const [hasFileRead, filePaths] = this.handlePotentialFileMentionCalls(
-									i,
-									secondBlock.text,
-									fileReadIndices,
-									thisExistingFileReads, // file reads we've already replaced in this text in the latest version of this updated text
-								)
-								if (hasFileRead) {
-									messageFilePaths.set(i, filePaths) // all file paths in this string
-								}
-							}
-						}
-					}
-				}
-			}
-		}
-
-		return [fileReadIndices, messageFilePaths]
-	}
-
-	/**
-	 * handles potential file content mentions in text blocks
-	 * there will not be more than one of the same file read in a text block
-	 */
-	private handlePotentialFileMentionCalls(
-		i: number,
-		secondBlockText: string,
-		fileReadIndices: Map<string, [number, number, string, string][]>,
-		thisExistingFileReads: string[],
-	): [boolean, string[]] {
-		const pattern = new RegExp(`<file_content path="([^"]*)">([\\s\\S]*?)</file_content>`, "g")
-
-		let foundMatch = false
-		const filePaths: string[] = []
-
-		let match
-		while ((match = pattern.exec(secondBlockText)) !== null) {
-			foundMatch = true
-
-			const filePath = match[1]
-			filePaths.push(filePath) // we will record all unique paths from file mentions in this text
-
-			// we can assume that thisExistingFileReads does not have many entries
-			if (!thisExistingFileReads.includes(filePath)) {
-				// meaning we havent already replaced this file read
-
-				const entireMatch = match[0] // The entire matched string
-
-				// Create the replacement text - keep the tags but replace the content
-				const replacementText = `<file_content path="${filePath}">${formatResponse.duplicateFileReadNotice()}</file_content>`
-
-				const indices = fileReadIndices.get(filePath) || []
-				indices.push([i, EditType.FILE_MENTION, entireMatch, replacementText])
-				fileReadIndices.set(filePath, indices)
-			}
-		}
-
-		return [foundMatch, filePaths]
-	}
-
-	/**
-	 * parses specific tool call formats, returns null if no acceptable format is found
-	 */
-	private parsePotentialToolCall(text: string): [string, string] | null {
-		const match = text.match(/^\[([^\s]+) for '([^']+)'\] Result:$/)
-
-		if (!match) {
-			return null
-		}
-
-		return [match[1], match[2]]
-	}
-
-	/**
-	 * file_read tool call always pastes the file, so this is always a hit
-	 */
-	private handleReadFileToolCall(
-		i: number,
-		filePath: string,
-		fileReadIndices: Map<string, [number, number, string, string][]>,
-	) {
-		const indices = fileReadIndices.get(filePath) || []
-		indices.push([i, EditType.READ_FILE_TOOL, "", formatResponse.duplicateFileReadNotice()])
-		fileReadIndices.set(filePath, indices)
-	}
-
-	/**
-	 * write_to_file and replace_in_file tool output are handled similarly
-	 */
-	private handlePotentialFileChangeToolCalls(
-		i: number,
-		filePath: string,
-		secondBlockText: string,
-		fileReadIndices: Map<string, [number, number, string, string][]>,
-	) {
-		const pattern = new RegExp(`(<final_file_content path="[^"]*">)[\\s\\S]*?(</final_file_content>)`)
-
-		// check if this exists in the text, it wont exist if the user rejects the file change for example
-		if (pattern.test(secondBlockText)) {
-			const replacementText = secondBlockText.replace(pattern, `$1 ${formatResponse.duplicateFileReadNotice()} $2`)
-			const indices = fileReadIndices.get(filePath) || []
-			indices.push([i, EditType.ALTER_FILE_TOOL, "", replacementText])
-			fileReadIndices.set(filePath, indices)
-		}
-	}
-
-	/**
-	 * alter all occurrences of file read operations and track which messages were updated
-	 * returns the outer index of messages we alter, to count number of changes
-	 */
-	private applyFileReadContextHistoryUpdates(
-		fileReadIndices: Map<string, [number, number, string, string][]>,
-		messageFilePaths: Map<number, string[]>,
-		apiMessages: Anthropic.Messages.MessageParam[],
-		timestamp: number,
-	): [boolean, Set<number>] {
-		let didUpdate = false
-		const updatedMessageIndices = new Set<number>() // track which messages we update on this round
-		const fileMentionUpdates = new Map<number, [string, string[]]>()
-
-		for (const [filePath, indices] of fileReadIndices.entries()) {
-			// Only process if there are multiple reads of the same file, else we will want to keep the latest read of the file
-			if (indices.length > 1) {
-				// Process all but the last index, as we will keep that instance of the file read
-				for (let i = 0; i < indices.length - 1; i++) {
-					const messageIndex = indices[i][0]
-					const messageType = indices[i][1] // EditType value
-					const searchText = indices[i][2] // search text (for file mentions, else empty string)
-					const messageString = indices[i][3] // what we will replace the string with
-
-					didUpdate = true
-					updatedMessageIndices.add(messageIndex)
-
-					// for single-fileread text we can set the updates here
-					// for potential multi-fileread text we need to determine all changes & iteratively update the text prior to saving the final change
-					if (messageType === EditType.FILE_MENTION) {
-						if (!fileMentionUpdates.has(messageIndex)) {
-							// Get base text either from existing updates or from apiMessages
-							let baseText = ""
-							let prevFilesReplaced: string[] = []
-
-							const innerTuple = this.contextHistoryUpdates.get(messageIndex)
-							if (innerTuple) {
-								const blockUpdates = innerTuple[1].get(1) // assumed index=1 for file mention filereads
-								if (blockUpdates && blockUpdates.length > 0) {
-									baseText = blockUpdates[blockUpdates.length - 1][2][0] // index 0 of MessageContent
-									prevFilesReplaced = blockUpdates[blockUpdates.length - 1][3][0] // previously overwritten file reads in this text
-								}
-							}
-
-							// can assume that this content will exist, otherwise it would not have been in fileReadIndices
-							const messageContent = apiMessages[messageIndex]?.content
-							if (!baseText && Array.isArray(messageContent) && messageContent.length > 1) {
-								const contentBlock = messageContent[1] // assume index=1 for all text to replace for file mention filereads
-								if (contentBlock.type === "text") {
-									baseText = contentBlock.text
-								}
-							}
-
-							// prevFilesReplaced keeps track of the previous file reads we've replace in this string, empty array if none
-							fileMentionUpdates.set(messageIndex, [baseText, prevFilesReplaced])
-						}
-
-						// Replace searchText with messageString for all file reads we need to replace in this text
-						if (searchText) {
-							const currentTuple = fileMentionUpdates.get(messageIndex) || ["", []]
-							if (currentTuple[0]) {
-								// safety check
-								// replace this text chunk
-								const updatedText = currentTuple[0].replace(searchText, messageString)
-
-								// add the newly added filePath read
-								const updatedFileReads = currentTuple[1]
-								updatedFileReads.push(filePath)
-
-								fileMentionUpdates.set(messageIndex, [updatedText, updatedFileReads])
-							}
-						}
-					} else {
-						let innerTuple = this.contextHistoryUpdates.get(messageIndex)
-						let innerMap: Map<number, ContextUpdate[]>
-
-						if (!innerTuple) {
-							innerMap = new Map<number, ContextUpdate[]>()
-							this.contextHistoryUpdates.set(messageIndex, [messageType, innerMap])
-						} else {
-							innerMap = innerTuple[1]
-						}
-
-						// block index for file reads from read_file, write_to_file, replace_in_file tools is 1
-						const blockIndex = 1
-
-						const updates = innerMap.get(blockIndex) || []
-
-						// metadata array is empty for non-file mention occurrences
-						updates.push([timestamp, "text", [messageString], []])
-
-						innerMap.set(blockIndex, updates)
-					}
-				}
-			}
-		}
-
-		// apply file mention updates to contextHistoryUpdates
-		// in fileMentionUpdates, filePathsUpdated includes all the file paths which are updated in the latest version of this altered text
-		for (const [messageIndex, [updatedText, filePathsUpdated]] of fileMentionUpdates.entries()) {
-			let innerTuple = this.contextHistoryUpdates.get(messageIndex)
-			let innerMap: Map<number, ContextUpdate[]>
-
-			if (!innerTuple) {
-				innerMap = new Map<number, ContextUpdate[]>()
-				this.contextHistoryUpdates.set(messageIndex, [EditType.FILE_MENTION, innerMap])
-			} else {
-				innerMap = innerTuple[1]
-			}
-
-			const blockIndex = 1 // we only consider the block index of 1 for file mentions
-			const updates = innerMap.get(blockIndex) || []
-
-			// filePathsUpdated includes changes done previously to this timestamp, and right now
-			if (messageFilePaths.has(messageIndex)) {
-				const allFileReads = messageFilePaths.get(messageIndex)
-				if (allFileReads) {
-					// safety check
-					// we gather all the file reads possible in this text from messageFilePaths
-					// filePathsUpdated from fileMentionUpdates stores all the files reads we have replaced now & previously
-					updates.push([timestamp, "text", [updatedText], [filePathsUpdated, allFileReads]])
-					innerMap.set(blockIndex, updates)
-				}
-			}
-		}
-
-		return [didUpdate, updatedMessageIndices]
-	}
-
-	/**
-	 * count total characters in messages and total savings within this range
-	 */
-	private countCharactersAndSavingsInRange(
-		apiMessages: Anthropic.Messages.MessageParam[],
-		startIndex: number,
-		endIndex: number,
-		uniqueFileReadIndices: Set<number>,
-	): { totalCharacters: number; charactersSaved: number } {
-		let totalCharCount = 0
-		let totalCharactersSaved = 0
-
-		for (let i = startIndex; i < endIndex; i++) {
-			// looping over the outer indicies of messages
-			const message = apiMessages[i]
-
-			if (!message.content) {
-				continue
-			}
-
-			// hasExistingAlterations checks whether the outer idnex has any changes
-			// hasExistingAlterations will also include the alterations we just made
-			const hasExistingAlterations = this.contextHistoryUpdates.has(i)
-			const hasNewAlterations = uniqueFileReadIndices.has(i)
-
-			if (Array.isArray(message.content)) {
-				for (let blockIndex = 0; blockIndex < message.content.length; blockIndex++) {
-					// looping over inner indices of messages
-					const block = message.content[blockIndex]
-
-					if (block.type === "text" && block.text) {
-						// true if we just altered it, or it was altered before
-						if (hasExistingAlterations) {
-							const innerTuple = this.contextHistoryUpdates.get(i)
-							const updates = innerTuple?.[1].get(blockIndex) // updated text for this inner index
-
-							if (updates && updates.length > 0) {
-								// exists if we have an update for the message at this index
-								const latestUpdate = updates[updates.length - 1]
-
-								// if block was just altered, then calculate savings
-								if (hasNewAlterations) {
-									let originalTextLength
-									if (updates.length > 1) {
-										originalTextLength = updates[updates.length - 2][2][0].length // handles case if we have multiple updates for same text block
-									} else {
-										originalTextLength = block.text.length
-									}
-
-									const newTextLength = latestUpdate[2][0].length // replacement text
-									totalCharactersSaved += originalTextLength - newTextLength
-
-									totalCharCount += originalTextLength
-								} else {
-									// meaning there was an update to this text previously, but we didnt just alter it
-									totalCharCount += latestUpdate[2][0].length
-								}
-							} else {
-								// reach here if there was one inner index with an update, but now we are at a different index, so updates is not defined
-								totalCharCount += block.text.length
-							}
-						} else {
-							// reach here if there's no alterations for this outer index, meaning each inner index wont have any changes either
-							totalCharCount += block.text.length
-						}
-					} else if (block.type === "image" && block.source) {
-						if (block.source.type === "base64" && block.source.data) {
-							totalCharCount += block.source.data.length
-						}
-					}
-				}
-			}
-		}
-
-		return { totalCharacters: totalCharCount, charactersSaved: totalCharactersSaved }
-	}
-
-	/**
-	 * count total percentage character savings across in-range conversation
-	 */
-	private calculateContextOptimizationMetrics(
-		apiMessages: Anthropic.Messages.MessageParam[],
-		conversationHistoryDeletedRange: [number, number] | undefined,
-		uniqueFileReadIndices: Set<number>,
-	): number {
-		// count for first user-assistant message pair
-		const firstChunkResult = this.countCharactersAndSavingsInRange(apiMessages, 0, 2, uniqueFileReadIndices)
-
-		// count for the remaining in-range messages
-		const secondChunkResult = this.countCharactersAndSavingsInRange(
-			apiMessages,
-			conversationHistoryDeletedRange ? conversationHistoryDeletedRange[1] + 1 : 2,
-			apiMessages.length,
-			uniqueFileReadIndices,
-		)
-
-		const totalCharacters = firstChunkResult.totalCharacters + secondChunkResult.totalCharacters
-		const totalCharactersSaved = firstChunkResult.charactersSaved + secondChunkResult.charactersSaved
-
-		const percentCharactersSaved = totalCharacters === 0 ? 0 : totalCharactersSaved / totalCharacters
-
-		return percentCharactersSaved
-	}
-}

src/core/context/context-management/ContextManager.ts

@@ -1,15 +1,120 @@
-import { Anthropic } from "@anthropic-ai/sdk"
+import { getContextWindowInfo } from "./context-window-utils"
+import { formatResponse } from "../../prompts/responses"
+import { GlobalFileNames } from "../../storage/disk"
+import { fileExistsAtPath } from "../../../utils/fs"
+import * as path from "path"
+import fs from "fs/promises"
+import cloneDeep from "clone-deep"
 import { ClineApiReqInfo, ClineMessage } from "../../../shared/ExtensionMessage"
 import { ApiHandler } from "../../../api"
-import { getContextWindowInfo } from "./context-window-utils"
+import { Anthropic } from "@anthropic-ai/sdk"
+
+enum EditType {
+	UNDEFINED = 0,
+	NO_FILE_READ = 1,
+	READ_FILE_TOOL = 2,
+	ALTER_FILE_TOOL = 3,
+	FILE_MENTION = 4,
+}
+
+// array of string values allows us to cover all changes for message types currently supported
+type MessageContent = string[]
+type MessageMetadata = string[][]
+
+// Type for a single context update
+type ContextUpdate = [number, string, MessageContent, MessageMetadata] // [timestamp, updateType, update, metadata]
+
+// Type for the serialized format of our nested maps
+type SerializedContextHistory = Array<
+	[
+		number, // messageIndex
+		[
+			number, // EditType (message type)
+			Array<
+				[
+					number, // blockIndex
+					ContextUpdate[], // updates array (now with 4 elements including metadata)
+				]
+			>,
+		],
+	]
+>
 
 export class ContextManager {
-	getNewContextMessagesAndMetadata(
+	// mapping from the apiMessages outer index to the inner message index to a list of actual changes, ordered by timestamp
+	// timestamp is required in order to support full checkpointing, where the changes we apply need to be able to be undone when
+	// moving to an earlier conversation history checkpoint - this ordering intuitively allows for binary search on truncation
+	// there is also a number stored for each (EditType) which defines which message type it is, for custom handling
+
+	// format:  { outerIndex => [EditType, { innerIndex => [[timestamp, updateType, update], ...] }] }
+	// example: { 1 => { [0, 0 => [[<timestamp>, "text", "[NOTE] Some previous conversation history with the user has been removed ..."], ...] }] }
+	// the above example would be how we update the first assistant message to indicate we truncated text
+	private contextHistoryUpdates: Map<number, [number, Map<number, ContextUpdate[]>]>
+
+	constructor() {
+		this.contextHistoryUpdates = new Map()
+	}
+
+	/**
+	 * public function for loading contextHistoryUpdates from disk, if it exists
+	 */
+	async initializeContextHistory(taskDirectory: string) {
+		this.contextHistoryUpdates = await this.getSavedContextHistory(taskDirectory)
+	}
+
+	/**
+	 * get the stored context history updates from disk
+	 */
+	private async getSavedContextHistory(taskDirectory: string): Promise<Map<number, [number, Map<number, ContextUpdate[]>]>> {
+		try {
+			const filePath = path.join(taskDirectory, GlobalFileNames.contextHistory)
+			if (await fileExistsAtPath(filePath)) {
+				const data = await fs.readFile(filePath, "utf8")
+				const serializedUpdates = JSON.parse(data) as SerializedContextHistory
+
+				// Update to properly reconstruct the tuple structure
+				return new Map(
+					serializedUpdates.map(([messageIndex, [numberValue, innerMapArray]]) => [
+						messageIndex,
+						[numberValue, new Map(innerMapArray)],
+					]),
+				)
+			}
+		} catch (error) {
+			console.error("Failed to load context history:", error)
+		}
+		return new Map()
+	}
+
+	/**
+	 * save the context history updates to disk
+	 */
+	private async saveContextHistory(taskDirectory: string) {
+		try {
+			const serializedUpdates: SerializedContextHistory = Array.from(this.contextHistoryUpdates.entries()).map(
+				([messageIndex, [numberValue, innerMap]]) => [messageIndex, [numberValue, Array.from(innerMap.entries())]],
+			)
+
+			await fs.writeFile(
+				path.join(taskDirectory, GlobalFileNames.contextHistory),
+				JSON.stringify(serializedUpdates),
+				"utf8",
+			)
+		} catch (error) {
+			console.error("Failed to save context history:", error)
+		}
+	}
+
+	/**
+	 * primary entry point for getting up to date context & truncating when required
+	 */
+	async getNewContextMessagesAndMetadata(
 		apiConversationHistory: Anthropic.Messages.MessageParam[],
 		clineMessages: ClineMessage[],
 		api: ApiHandler,
 		conversationHistoryDeletedRange: [number, number] | undefined,
 		previousApiReqIndex: number,
+		taskDirectory: string,
 	) {
 		let updatedConversationHistoryDeletedRange = false
 
@@ -17,6 +122,7 @@ export class ContextManager {
 		if (previousApiReqIndex >= 0) {
 			const previousRequest = clineMessages[previousApiReqIndex]
 			if (previousRequest && previousRequest.text) {
+				const timestamp = previousRequest.ts
 				const { tokensIn, tokensOut, cacheWrites, cacheReads }: ClineApiReqInfo = JSON.parse(previousRequest.text)
 				const totalTokens = (tokensIn || 0) + (tokensOut || 0) + (cacheWrites || 0) + (cacheReads || 0)
 				const { maxAllowedSize } = getContextWindowInfo(api)
@@ -25,23 +131,54 @@ export class ContextManager {
 				if (totalTokens >= maxAllowedSize) {
 					// Since the user may switch between models with different context windows, truncating half may not be enough (ie if switching from claude 200k to deepseek 64k, half truncation will only remove 100k tokens, but we need to remove much more)
 					// So if totalTokens/2 is greater than maxAllowedSize, we truncate 3/4 instead of 1/2
-					// FIXME: truncating the conversation in a way that is optimal for prompt caching AND takes into account multi-context window complexity is something we need to improve
 					const keep = totalTokens / 2 > maxAllowedSize ? "quarter" : "half"
 
-					// NOTE: it's okay that we overwriteConversationHistory in resume task since we're only ever removing the last user message and not anything in the middle which would affect this range
-					conversationHistoryDeletedRange = this.getNextTruncationRange(
+					// we later check how many chars we trim to determine if we should still truncate history
+					let [anyContextUpdates, uniqueFileReadIndices] = this.applyContextOptimizations(
 						apiConversationHistory,
-						conversationHistoryDeletedRange,
-						keep,
+						conversationHistoryDeletedRange ? conversationHistoryDeletedRange[1] + 1 : 2,
+						timestamp,
 					)
 
-					updatedConversationHistoryDeletedRange = true
+					let needToTruncate = true
+					if (anyContextUpdates) {
+						// determine whether we've saved enough chars to not truncate
+						const charactersSavedPercentage = this.calculateContextOptimizationMetrics(
+							apiConversationHistory,
+							conversationHistoryDeletedRange,
+							uniqueFileReadIndices,
+						)
+						if (charactersSavedPercentage >= 0.3) {
+							needToTruncate = false
+						}
+					}
+
+					if (needToTruncate) {
+						// go ahead with truncation
+						anyContextUpdates = this.applyStandardContextTruncationNoticeChange(timestamp) || anyContextUpdates
+
+						// NOTE: it's okay that we overwriteConversationHistory in resume task since we're only ever removing the last user message and not anything in the middle which would affect this range
+						conversationHistoryDeletedRange = this.getNextTruncationRange(
+							apiConversationHistory,
+							conversationHistoryDeletedRange,
+							keep,
+						)
+
+						updatedConversationHistoryDeletedRange = true
+					}
+
+					// if we alter the context history, save the updated version to disk
+					if (anyContextUpdates) {
+						await this.saveContextHistory(taskDirectory)
+					}
 				}
 			}
 		}
 
-		// conversationHistoryDeletedRange is updated only when we're close to hitting the context window, so we don't continuously break the prompt cache
-		const truncatedConversationHistory = this.getTruncatedMessages(apiConversationHistory, conversationHistoryDeletedRange)
+		const truncatedConversationHistory = this.getAndAlterTruncatedMessages(
+			apiConversationHistory,
+			conversationHistoryDeletedRange,
+		)
 
 		return {
 			conversationHistoryDeletedRange: conversationHistoryDeletedRange,
@@ -50,14 +187,17 @@ export class ContextManager {
 		}
 	}
 
+	/**
+	 * get truncation range
+	 */
 	public getNextTruncationRange(
 		apiMessages: Anthropic.Messages.MessageParam[],
 		currentDeletedRange: [number, number] | undefined,
 		keep: "half" | "quarter",
 	): [number, number] {
-		// Since we always keep the first message, currentDeletedRange[0] will always be 1 (for now until we have a smarter truncation algorithm)
-		const rangeStartIndex = 1
-		const startOfRest = currentDeletedRange ? currentDeletedRange[1] + 1 : 1
+		// We always keep the first user-assistant pairing, and truncate an even number of messages from there
+		const rangeStartIndex = 2 // index 0 and 1 are kept
+		const startOfRest = currentDeletedRange ? currentDeletedRange[1] + 1 : 2 // inclusive starting index
 
 		let messagesToRemove: number
 		if (keep === "half") {
@@ -74,11 +214,11 @@ export class ContextManager {
 			messagesToRemove = Math.floor(((apiMessages.length - startOfRest) * 3) / 4 / 2) * 2
 		}
 
-		let rangeEndIndex = startOfRest + messagesToRemove - 1
+		let rangeEndIndex = startOfRest + messagesToRemove - 1 // inclusive ending index
 
-		// Make sure the last message being removed is a user message, so that the next message after the initial task message is an assistant message. This preservers the user-assistant-user-assistant structure.
+		// Make sure that the last message being removed is a assistant message, so the next message after the initial user-assistant pair is an assistant message. This preserves the user-assistant-user-assistant structure.
 		// NOTE: anthropic format messages are always user-assistant-user-assistant, while openai format messages can have multiple user messages in a row (we use anthropic format throughout cline)
-		if (apiMessages[rangeEndIndex].role !== "user") {
+		if (apiMessages[rangeEndIndex].role !== "assistant") {
 			rangeEndIndex -= 1
 		}
 
@@ -86,17 +226,609 @@ export class ContextManager {
 		return [rangeStartIndex, rangeEndIndex]
 	}
 
+	/**
+	 * external interface to support old calls
+	 */
 	public getTruncatedMessages(
 		messages: Anthropic.Messages.MessageParam[],
 		deletedRange: [number, number] | undefined,
 	): Anthropic.Messages.MessageParam[] {
-		if (!deletedRange) {
+		return this.getAndAlterTruncatedMessages(messages, deletedRange)
+	}
+
+	/**
+	 * apply all required truncation methods to the messages in context
+	 */
+	private getAndAlterTruncatedMessages(
+		messages: Anthropic.Messages.MessageParam[],
+		deletedRange: [number, number] | undefined,
+	): Anthropic.Messages.MessageParam[] {
+		if (messages.length <= 1) {
 			return messages
 		}
 
-		const [start, end] = deletedRange
-		// the range is inclusive - both start and end indices and everything in between will be removed from the final result.
-		// NOTE: if you try to console log these, don't forget that logging a reference to an array may not provide the same result as logging a slice() snapshot of that array at that exact moment. The following DOES in fact include the latest assistant message.
-		return [...messages.slice(0, start), ...messages.slice(end + 1)]
+		const updatedMessages = this.applyContextHistoryUpdates(messages, deletedRange ? deletedRange[1] + 1 : 2)
+
+		// OLD NOTE: if you try to console log these, don't forget that logging a reference to an array may not provide the same result as logging a slice() snapshot of that array at that exact moment. The following DOES in fact include the latest assistant message.
+		return updatedMessages
+	}
+
+	/**
+	 * applies deletedRange truncation and other alterations based on changes in this.contextHistoryUpdates
+	 */
+	private applyContextHistoryUpdates(
+		messages: Anthropic.Messages.MessageParam[],
+		startFromIndex: number,
+	): Anthropic.Messages.MessageParam[] {
+		// runtime is linear in length of user messages, if expecting a limited number of alterations, could be more optimal to loop over alterations
+
+		const firstChunk = messages.slice(0, 2) // get first user-assistant pair
+		const secondChunk = messages.slice(startFromIndex) // get remaining messages within context
+		const messagesToUpdate = [...firstChunk, ...secondChunk]
+
+		// we need the mapping from the local indices in messagesToUpdate to the global array of updates in this.contextHistoryUpdates
+		const originalIndices = [
+			...Array(2).keys(),
+			...Array(secondChunk.length)
+				.fill(0)
+				.map((_, i) => i + startFromIndex),
+		]
+
+		for (let arrayIndex = 0; arrayIndex < messagesToUpdate.length; arrayIndex++) {
+			const messageIndex = originalIndices[arrayIndex]
+
+			const innerTuple = this.contextHistoryUpdates.get(messageIndex)
+			if (!innerTuple) {
+				continue
+			}
+
+			// because we are altering this, we need a deep copy
+			messagesToUpdate[arrayIndex] = cloneDeep(messagesToUpdate[arrayIndex])
+
+			// Extract the map from the tuple
+			const innerMap = innerTuple[1]
+			for (const [blockIndex, changes] of innerMap) {
+				// apply the latest change among n changes - [timestamp, updateType, update]
+				const latestChange = changes[changes.length - 1]
+
+				if (latestChange[1] === "text") {
+					// only altering text for now
+					const message = messagesToUpdate[arrayIndex]
+
+					if (Array.isArray(message.content)) {
+						const block = message.content[blockIndex]
+						if (block && block.type === "text") {
+							block.text = latestChange[2][0]
+						}
+					}
+				}
+			}
+		}
+
+		return messagesToUpdate
+	}
+
+	/**
+	 * removes all context history updates that occurred after the specified timestamp and saves to disk
+	 */
+	async truncateContextHistory(timestamp: number, taskDirectory: string): Promise<void> {
+		this.truncateContextHistoryAtTimestamp(this.contextHistoryUpdates, timestamp)
+
+		// save the modified context history to disk
+		await this.saveContextHistory(taskDirectory)
+	}
+
+	/**
+	 * alters the context history to remove all alterations after a given timestamp
+	 * removes the index if there are no alterations there anymore, both outer and inner indices
+	 */
+	private truncateContextHistoryAtTimestamp(
+		contextHistory: Map<number, [number, Map<number, ContextUpdate[]>]>,
+		timestamp: number,
+	): void {
+		for (const [messageIndex, [_, innerMap]] of contextHistory) {
+			// track which blockIndices to delete
+			const blockIndicesToDelete: number[] = []
+
+			// loop over the innerIndices of the messages in this block
+			for (const [blockIndex, updates] of innerMap) {
+				// updates ordered by timestamp, so find cutoff point by iterating from right to left
+				let cutoffIndex = updates.length - 1
+				while (cutoffIndex >= 0 && updates[cutoffIndex][0] > timestamp) {
+					cutoffIndex--
+				}
+
+				// If we found updates to remove
+				if (cutoffIndex < updates.length - 1) {
+					// Modify the array in place to keep only updates up to cutoffIndex
+					updates.length = cutoffIndex + 1
+
+					// If no updates left after truncation, mark this block for deletion
+					if (updates.length === 0) {
+						blockIndicesToDelete.push(blockIndex)
+					}
+				}
+			}
+
+			// Remove empty blocks from inner map
+			for (const blockIndex of blockIndicesToDelete) {
+				innerMap.delete(blockIndex)
+			}
+
+			// If inner map is now empty, remove the message index from outer map
+			if (innerMap.size === 0) {
+				contextHistory.delete(messageIndex)
+			}
+		}
+	}
+
+	/**
+	 * applies the context optimization steps and returns whether any changes were made
+	 */
+	private applyContextOptimizations(
+		apiMessages: Anthropic.Messages.MessageParam[],
+		startFromIndex: number,
+		timestamp: number,
+	): [boolean, Set<number>] {
+		const [fileReadUpdatesBool, uniqueFileReadIndices] = this.findAndPotentiallySaveFileReadContextHistoryUpdates(
+			apiMessages,
+			startFromIndex,
+			timestamp,
+		)
+
+		// true if any context optimization steps alter state
+		const contextHistoryUpdated = fileReadUpdatesBool
+
+		return [contextHistoryUpdated, uniqueFileReadIndices]
+	}
+
+	/**
+	 * if there is any truncation and there is no other alteration already set, alter the assistant message to indicate this occurred
+	 */
+	private applyStandardContextTruncationNoticeChange(timestamp: number): boolean {
+		if (!this.contextHistoryUpdates.has(1)) {
+			// first assistant message always at index 1
+			const innerMap = new Map<number, ContextUpdate[]>()
+			innerMap.set(0, [[timestamp, "text", [formatResponse.contextTruncationNotice()], []]])
+			this.contextHistoryUpdates.set(1, [0, innerMap]) // EditType is undefined for first assistant message
+			return true
+		}
+		return false
+	}
+
+	/**
+	 * wraps the logic for determining file reads to overwrite, and altering state
+	 * returns whether any updates were made (bool) and indices where updates were made
+	 */
+	private findAndPotentiallySaveFileReadContextHistoryUpdates(
+		apiMessages: Anthropic.Messages.MessageParam[],
+		startFromIndex: number,
+		timestamp: number,
+	): [boolean, Set<number>] {
+		const [fileReadIndices, messageFilePaths] = this.getPossibleDuplicateFileReads(apiMessages, startFromIndex)
+		return this.applyFileReadContextHistoryUpdates(fileReadIndices, messageFilePaths, apiMessages, timestamp)
+	}
+
+	/**
+	 * generate a mapping from unique file reads from multiple tool calls to their outer index position(s)
+	 * also return additional metadata to support multiple file reads in file mention text blocks
+	 */
+	private getPossibleDuplicateFileReads(
+		apiMessages: Anthropic.Messages.MessageParam[],
+		startFromIndex: number,
+	): [Map<string, [number, number, string, string][]>, Map<number, string[]>] {
+		// fileReadIndices: { fileName => [outerIndex, EditType, searchText, replaceText] }
+		// messageFilePaths: { outerIndex => [fileRead1, fileRead2, ..] }
+		// searchText in fileReadIndices is only required for file mention file-reads since there can be more than one file in the text
+		// searchText will be the empty string "" in the case that it's not required, for non-file mentions
+		// messageFilePaths is only used for file mentions as there can be multiple files read in the same text chunk
+
+		// for all text blocks per file, has info for updating the block
+		const fileReadIndices = new Map<string, [number, number, string, string][]>()
+
+		// for file mention text blocks, track all the unique files read
+		const messageFilePaths = new Map<number, string[]>()
+
+		for (let i = startFromIndex; i < apiMessages.length; i++) {
+			let thisExistingFileReads: string[] = []
+
+			if (this.contextHistoryUpdates.has(i)) {
+				const innerTuple = this.contextHistoryUpdates.get(i)
+
+				if (innerTuple) {
+					// safety check
+					const editType = innerTuple[0]
+
+					if (editType === EditType.FILE_MENTION) {
+						const innerMap = innerTuple[1]
+
+						const blockIndex = 1 // file mention blocks assumed to be at index 1
+						const blockUpdates = innerMap.get(blockIndex)
+
+						// if we have updated this text previously, we want to check whether the lists of files in the metadata are the same
+						if (blockUpdates && blockUpdates.length > 0) {
+							// the first list indicates the files we have replaced in this text, second list indicates all unique files in this text
+							// if they are equal then we have replaced all the files in this text already, and can ignore further processing
+							if (
+								blockUpdates[blockUpdates.length - 1][3][0].length ===
+								blockUpdates[blockUpdates.length - 1][3][1].length
+							) {
+								continue
+							}
+							// otherwise there are still file reads here we can overwrite, so still need to process this text chunk
+							// to do so we need to keep track of which files we've already replaced so we don't replace them again
+							else {
+								thisExistingFileReads = blockUpdates[blockUpdates.length - 1][3][0]
+							}
+						}
+					} else {
+						// for all other cases we can assume that we dont need to check this again
+						continue
+					}
+				}
+			}
+
+			const message = apiMessages[i]
+			if (message.role === "user" && Array.isArray(message.content) && message.content.length > 0) {
+				const firstBlock = message.content[0]
+				if (firstBlock.type === "text") {
+					const matchTup = this.parsePotentialToolCall(firstBlock.text)
+					let foundNormalFileRead = false
+					if (matchTup) {
+						if (matchTup[0] === "read_file") {
+							this.handleReadFileToolCall(i, matchTup[1], fileReadIndices)
+							foundNormalFileRead = true
+						} else if (matchTup[0] === "replace_in_file" || matchTup[0] === "write_to_file") {
+							if (message.content.length > 1) {
+								const secondBlock = message.content[1]
+								if (secondBlock.type === "text") {
+									this.handlePotentialFileChangeToolCalls(i, matchTup[1], secondBlock.text, fileReadIndices)
+									foundNormalFileRead = true
+								}
+							}
+						}
+					}
+
+					// file mentions can happen in most other user message blocks
+					if (!foundNormalFileRead) {
+						if (message.content.length > 1) {
+							const secondBlock = message.content[1]
+							if (secondBlock.type === "text") {
+								const [hasFileRead, filePaths] = this.handlePotentialFileMentionCalls(
+									i,
+									secondBlock.text,
+									fileReadIndices,
+									thisExistingFileReads, // file reads we've already replaced in this text in the latest version of this updated text
+								)
+								if (hasFileRead) {
+									messageFilePaths.set(i, filePaths) // all file paths in this string
+								}
+							}
+						}
+					}
+				}
+			}
+		}
+
+		return [fileReadIndices, messageFilePaths]
+	}
+
+	/**
+	 * handles potential file content mentions in text blocks
+	 * there will not be more than one of the same file read in a text block
+	 */
+	private handlePotentialFileMentionCalls(
+		i: number,
+		secondBlockText: string,
+		fileReadIndices: Map<string, [number, number, string, string][]>,
+		thisExistingFileReads: string[],
+	): [boolean, string[]] {
+		const pattern = new RegExp(`<file_content path="([^"]*)">([\\s\\S]*?)</file_content>`, "g")
+
+		let foundMatch = false
+		const filePaths: string[] = []
+
+		let match
+		while ((match = pattern.exec(secondBlockText)) !== null) {
+			foundMatch = true
+
+			const filePath = match[1]
+			filePaths.push(filePath) // we will record all unique paths from file mentions in this text
+
+			// we can assume that thisExistingFileReads does not have many entries
+			if (!thisExistingFileReads.includes(filePath)) {
+				// meaning we haven't already replaced this file read
+
+				const entireMatch = match[0] // The entire matched string
+
+				// Create the replacement text - keep the tags but replace the content
+				const replacementText = `<file_content path="${filePath}">${formatResponse.duplicateFileReadNotice()}</file_content>`
+
+				const indices = fileReadIndices.get(filePath) || []
+				indices.push([i, EditType.FILE_MENTION, entireMatch, replacementText])
+				fileReadIndices.set(filePath, indices)
+			}
+		}
+
+		return [foundMatch, filePaths]
+	}
+
+	/**
+	 * parses specific tool call formats, returns null if no acceptable format is found
+	 */
+	private parsePotentialToolCall(text: string): [string, string] | null {
+		const match = text.match(/^\[([^\s]+) for '([^']+)'\] Result:$/)
+
+		if (!match) {
+			return null
+		}
+
+		return [match[1], match[2]]
+	}
+
+	/**
+	 * file_read tool call always pastes the file, so this is always a hit
+	 */
+	private handleReadFileToolCall(
+		i: number,
+		filePath: string,
+		fileReadIndices: Map<string, [number, number, string, string][]>,
+	) {
+		const indices = fileReadIndices.get(filePath) || []
+		indices.push([i, EditType.READ_FILE_TOOL, "", formatResponse.duplicateFileReadNotice()])
+		fileReadIndices.set(filePath, indices)
+	}
+
+	/**
+	 * write_to_file and replace_in_file tool output are handled similarly
+	 */
+	private handlePotentialFileChangeToolCalls(
+		i: number,
+		filePath: string,
+		secondBlockText: string,
+		fileReadIndices: Map<string, [number, number, string, string][]>,
+	) {
+		const pattern = new RegExp(`(<final_file_content path="[^"]*">)[\\s\\S]*?(</final_file_content>)`)
+
+		// check if this exists in the text, it won't exist if the user rejects the file change for example
+		if (pattern.test(secondBlockText)) {
+			const replacementText = secondBlockText.replace(pattern, `$1 ${formatResponse.duplicateFileReadNotice()} $2`)
+			const indices = fileReadIndices.get(filePath) || []
+			indices.push([i, EditType.ALTER_FILE_TOOL, "", replacementText])
+			fileReadIndices.set(filePath, indices)
+		}
+	}
+
+	/**
+	 * alter all occurrences of file read operations and track which messages were updated
+	 * returns the outer index of messages we alter, to count number of changes
+	 */
+	private applyFileReadContextHistoryUpdates(
+		fileReadIndices: Map<string, [number, number, string, string][]>,
+		messageFilePaths: Map<number, string[]>,
+		apiMessages: Anthropic.Messages.MessageParam[],
+		timestamp: number,
+	): [boolean, Set<number>] {
+		let didUpdate = false
+		const updatedMessageIndices = new Set<number>() // track which messages we update on this round
+		const fileMentionUpdates = new Map<number, [string, string[]]>()
+
+		for (const [filePath, indices] of fileReadIndices.entries()) {
+			// Only process if there are multiple reads of the same file, else we will want to keep the latest read of the file
+			if (indices.length > 1) {
+				// Process all but the last index, as we will keep that instance of the file read
+				for (let i = 0; i < indices.length - 1; i++) {
+					const messageIndex = indices[i][0]
+					const messageType = indices[i][1] // EditType value
+					const searchText = indices[i][2] // search text (for file mentions, else empty string)
+					const messageString = indices[i][3] // what we will replace the string with
+
+					didUpdate = true
+					updatedMessageIndices.add(messageIndex)
+
+					// for single-fileread text we can set the updates here
+					// for potential multi-fileread text we need to determine all changes & iteratively update the text prior to saving the final change
+					if (messageType === EditType.FILE_MENTION) {
+						if (!fileMentionUpdates.has(messageIndex)) {
+							// Get base text either from existing updates or from apiMessages
+							let baseText = ""
+							let prevFilesReplaced: string[] = []
+
+							const innerTuple = this.contextHistoryUpdates.get(messageIndex)
+							if (innerTuple) {
+								const blockUpdates = innerTuple[1].get(1) // assumed index=1 for file mention filereads
+								if (blockUpdates && blockUpdates.length > 0) {
+									baseText = blockUpdates[blockUpdates.length - 1][2][0] // index 0 of MessageContent
+									prevFilesReplaced = blockUpdates[blockUpdates.length - 1][3][0] // previously overwritten file reads in this text
+								}
+							}
+
+							// can assume that this content will exist, otherwise it would not have been in fileReadIndices
+							const messageContent = apiMessages[messageIndex]?.content
+							if (!baseText && Array.isArray(messageContent) && messageContent.length > 1) {
+								const contentBlock = messageContent[1] // assume index=1 for all text to replace for file mention filereads
+								if (contentBlock.type === "text") {
+									baseText = contentBlock.text
+								}
+							}
+
+							// prevFilesReplaced keeps track of the previous file reads we've replace in this string, empty array if none
+							fileMentionUpdates.set(messageIndex, [baseText, prevFilesReplaced])
+						}
+
+						// Replace searchText with messageString for all file reads we need to replace in this text
+						if (searchText) {
+							const currentTuple = fileMentionUpdates.get(messageIndex) || ["", []]
+							if (currentTuple[0]) {
+								// safety check
+								// replace this text chunk
+								const updatedText = currentTuple[0].replace(searchText, messageString)
+
+								// add the newly added filePath read
+								const updatedFileReads = currentTuple[1]
+								updatedFileReads.push(filePath)
+
+								fileMentionUpdates.set(messageIndex, [updatedText, updatedFileReads])
+							}
+						}
+					} else {
+						let innerTuple = this.contextHistoryUpdates.get(messageIndex)
+						let innerMap: Map<number, ContextUpdate[]>
+
+						if (!innerTuple) {
+							innerMap = new Map<number, ContextUpdate[]>()
+							this.contextHistoryUpdates.set(messageIndex, [messageType, innerMap])
+						} else {
+							innerMap = innerTuple[1]
+						}
+
+						// block index for file reads from read_file, write_to_file, replace_in_file tools is 1
+						const blockIndex = 1
+
+						const updates = innerMap.get(blockIndex) || []
+
+						// metadata array is empty for non-file mention occurrences
+						updates.push([timestamp, "text", [messageString], []])
+
+						innerMap.set(blockIndex, updates)
+					}
+				}
+			}
+		}
+
+		// apply file mention updates to contextHistoryUpdates
+		// in fileMentionUpdates, filePathsUpdated includes all the file paths which are updated in the latest version of this altered text
+		for (const [messageIndex, [updatedText, filePathsUpdated]] of fileMentionUpdates.entries()) {
+			let innerTuple = this.contextHistoryUpdates.get(messageIndex)
+			let innerMap: Map<number, ContextUpdate[]>
+
+			if (!innerTuple) {
+				innerMap = new Map<number, ContextUpdate[]>()
+				this.contextHistoryUpdates.set(messageIndex, [EditType.FILE_MENTION, innerMap])
+			} else {
+				innerMap = innerTuple[1]
+			}
+
+			const blockIndex = 1 // we only consider the block index of 1 for file mentions
+			const updates = innerMap.get(blockIndex) || []
+
+			// filePathsUpdated includes changes done previously to this timestamp, and right now
+			if (messageFilePaths.has(messageIndex)) {
+				const allFileReads = messageFilePaths.get(messageIndex)
+				if (allFileReads) {
+					// safety check
+					// we gather all the file reads possible in this text from messageFilePaths
+					// filePathsUpdated from fileMentionUpdates stores all the files reads we have replaced now & previously
+					updates.push([timestamp, "text", [updatedText], [filePathsUpdated, allFileReads]])
+					innerMap.set(blockIndex, updates)
+				}
+			}
+		}
+
+		return [didUpdate, updatedMessageIndices]
+	}
+
+	/**
+	 * count total characters in messages and total savings within this range
+	 */
+	private countCharactersAndSavingsInRange(
+		apiMessages: Anthropic.Messages.MessageParam[],
+		startIndex: number,
+		endIndex: number,
+		uniqueFileReadIndices: Set<number>,
+	): { totalCharacters: number; charactersSaved: number } {
+		let totalCharCount = 0
+		let totalCharactersSaved = 0
+
+		for (let i = startIndex; i < endIndex; i++) {
+			// looping over the outer indices of messages
+			const message = apiMessages[i]
+
+			if (!message.content) {
+				continue
+			}
+
+			// hasExistingAlterations checks whether the outer idnex has any changes
+			// hasExistingAlterations will also include the alterations we just made
+			const hasExistingAlterations = this.contextHistoryUpdates.has(i)
+			const hasNewAlterations = uniqueFileReadIndices.has(i)
+
+			if (Array.isArray(message.content)) {
+				for (let blockIndex = 0; blockIndex < message.content.length; blockIndex++) {
+					// looping over inner indices of messages
+					const block = message.content[blockIndex]
+
+					if (block.type === "text" && block.text) {
+						// true if we just altered it, or it was altered before
+						if (hasExistingAlterations) {
+							const innerTuple = this.contextHistoryUpdates.get(i)
+							const updates = innerTuple?.[1].get(blockIndex) // updated text for this inner index
+
+							if (updates && updates.length > 0) {
+								// exists if we have an update for the message at this index
+								const latestUpdate = updates[updates.length - 1]
+
+								// if block was just altered, then calculate savings
+								if (hasNewAlterations) {
+									let originalTextLength
+									if (updates.length > 1) {
+										originalTextLength = updates[updates.length - 2][2][0].length // handles case if we have multiple updates for same text block
+									} else {
+										originalTextLength = block.text.length
+									}
+
+									const newTextLength = latestUpdate[2][0].length // replacement text
+									totalCharactersSaved += originalTextLength - newTextLength
+
+									totalCharCount += originalTextLength
+								} else {
+									// meaning there was an update to this text previously, but we didn't just alter it
+									totalCharCount += latestUpdate[2][0].length
+								}
+							} else {
+								// reach here if there was one inner index with an update, but now we are at a different index, so updates is not defined
+								totalCharCount += block.text.length
+							}
+						} else {
+							// reach here if there's no alterations for this outer index, meaning each inner index won't have any changes either
+							totalCharCount += block.text.length
+						}
+					} else if (block.type === "image" && block.source) {
+						if (block.source.type === "base64" && block.source.data) {
+							totalCharCount += block.source.data.length
+						}
+					}
+				}
+			}
+		}
+
+		return { totalCharacters: totalCharCount, charactersSaved: totalCharactersSaved }
+	}
+
+	/**
+	 * count total percentage character savings across in-range conversation
+	 */
+	private calculateContextOptimizationMetrics(
+		apiMessages: Anthropic.Messages.MessageParam[],
+		conversationHistoryDeletedRange: [number, number] | undefined,
+		uniqueFileReadIndices: Set<number>,
+	): number {
+		// count for first user-assistant message pair
+		const firstChunkResult = this.countCharactersAndSavingsInRange(apiMessages, 0, 2, uniqueFileReadIndices)
+
+		// count for the remaining in-range messages
+		const secondChunkResult = this.countCharactersAndSavingsInRange(
+			apiMessages,
+			conversationHistoryDeletedRange ? conversationHistoryDeletedRange[1] + 1 : 2,
+			apiMessages.length,
+			uniqueFileReadIndices,
+		)
+
+		const totalCharacters = firstChunkResult.totalCharacters + secondChunkResult.totalCharacters
+		const totalCharactersSaved = firstChunkResult.charactersSaved + secondChunkResult.charactersSaved
+
+		const percentCharactersSaved = totalCharacters === 0 ? 0 : totalCharactersSaved / totalCharacters
+
+		return percentCharactersSaved
 	}
 }

src/core/context/instructions/user-instructions/cline-rules.ts

@@ -1,10 +1,42 @@
 import path from "path"
-import { GlobalFileNames } from "../../../storage/disk"
+import { ensureRulesDirectoryExists, GlobalFileNames } from "../../../storage/disk"
 import { fileExistsAtPath, isDirectory, readDirectory } from "../../../../utils/fs"
 import { formatResponse } from "../../../prompts/responses"
 import fs from "fs/promises"
+import { ClineRulesToggles } from "../../../../shared/cline-rules"
+import { getGlobalState, getWorkspaceState, updateGlobalState, updateWorkspaceState } from "../../../storage/state"
+import * as vscode from "vscode"
 
-export const getClineRules = async (cwd: string) => {
+export const getGlobalClineRules = async (globalClineRulesFilePath: string, toggles: ClineRulesToggles) => {
+	if (await fileExistsAtPath(globalClineRulesFilePath)) {
+		if (await isDirectory(globalClineRulesFilePath)) {
+			try {
+				const rulesFilePaths = await readDirectory(globalClineRulesFilePath)
+				const rulesFilesTotalContent = await getClineRulesFilesTotalContent(
+					rulesFilePaths,
+					globalClineRulesFilePath,
+					toggles,
+				)
+				if (rulesFilesTotalContent) {
+					const clineRulesFileInstructions = formatResponse.clineRulesGlobalDirectoryInstructions(
+						globalClineRulesFilePath,
+						rulesFilesTotalContent,
+					)
+					return clineRulesFileInstructions
+				}
+			} catch {
+				console.error(`Failed to read .clinerules directory at ${globalClineRulesFilePath}`)
+			}
+		} else {
+			console.error(`${globalClineRulesFilePath} is not a directory`)
+			return undefined
+		}
+	}
+
+	return undefined
+}
+
+export const getLocalClineRules = async (cwd: string, toggles: ClineRulesToggles) => {
 	const clineRulesFilePath = path.resolve(cwd, GlobalFileNames.clineRules)
 
 	let clineRulesFileInstructions: string | undefined
@@ -12,17 +44,21 @@ export const getClineRules = async (cwd: string) => {
 	if (await fileExistsAtPath(clineRulesFilePath)) {
 		if (await isDirectory(clineRulesFilePath)) {
 			try {
-				const rulesFilePaths = await readDirectory(path.join(cwd, GlobalFileNames.clineRules))
-				const rulesFilesTotalContent = await getClineRulesFilesTotalContent(rulesFilePaths, cwd)
-				clineRulesFileInstructions = formatResponse.clineRulesDirectoryInstructions(cwd, rulesFilesTotalContent)
+				const rulesFilePaths = await readDirectory(clineRulesFilePath)
+				const rulesFilesTotalContent = await getClineRulesFilesTotalContent(rulesFilePaths, cwd, toggles)
+				if (rulesFilesTotalContent) {
+					clineRulesFileInstructions = formatResponse.clineRulesLocalDirectoryInstructions(cwd, rulesFilesTotalContent)
+				}
 			} catch {
 				console.error(`Failed to read .clinerules directory at ${clineRulesFilePath}`)
 			}
 		} else {
 			try {
-				const ruleFileContent = (await fs.readFile(clineRulesFilePath, "utf8")).trim()
-				if (ruleFileContent) {
-					clineRulesFileInstructions = formatResponse.clineRulesFileInstructions(cwd, ruleFileContent)
+				if (clineRulesFilePath in toggles && toggles[clineRulesFilePath] !== false) {
+					const ruleFileContent = (await fs.readFile(clineRulesFilePath, "utf8")).trim()
+					if (ruleFileContent) {
+						clineRulesFileInstructions = formatResponse.clineRulesLocalFileInstructions(cwd, ruleFileContent)
+					}
 				}
 			} catch {
 				console.error(`Failed to read .clinerules file at ${clineRulesFilePath}`)
@@ -33,13 +69,107 @@ export const getClineRules = async (cwd: string) => {
 	return clineRulesFileInstructions
 }
 
-const getClineRulesFilesTotalContent = async (rulesFilePaths: string[], cwd: string) => {
+const getClineRulesFilesTotalContent = async (rulesFilePaths: string[], basePath: string, toggles: ClineRulesToggles) => {
 	const ruleFilesTotalContent = await Promise.all(
 		rulesFilePaths.map(async (filePath) => {
-			const ruleFilePath = path.resolve(cwd, filePath)
-			const ruleFilePathRelative = path.relative(cwd, ruleFilePath)
+			const ruleFilePath = path.resolve(basePath, filePath)
+			const ruleFilePathRelative = path.relative(basePath, ruleFilePath)
+
+			if (ruleFilePath in toggles && toggles[ruleFilePath] === false) {
+				return null
+			}
+
 			return `${ruleFilePathRelative}\n` + (await fs.readFile(ruleFilePath, "utf8")).trim()
 		}),
-	).then((contents) => contents.join("\n\n"))
+	).then((contents) => contents.filter(Boolean).join("\n\n"))
 	return ruleFilesTotalContent
 }
+
+export async function synchronizeRuleToggles(
+	rulesDirectoryPath: string,
+	currentToggles: ClineRulesToggles,
+): Promise<ClineRulesToggles> {
+	// Create a copy of toggles to modify
+	const updatedToggles = { ...currentToggles }
+
+	try {
+		const pathExists = await fileExistsAtPath(rulesDirectoryPath)
+
+		if (pathExists) {
+			const isDir = await isDirectory(rulesDirectoryPath)
+
+			if (isDir) {
+				// DIRECTORY CASE
+				const filePaths = await readDirectory(rulesDirectoryPath)
+				const existingRulePaths = new Set<string>()
+
+				for (const filePath of filePaths) {
+					const ruleFilePath = path.resolve(rulesDirectoryPath, filePath)
+					existingRulePaths.add(ruleFilePath)
+
+					const pathHasToggle = ruleFilePath in updatedToggles
+					if (!pathHasToggle) {
+						updatedToggles[ruleFilePath] = true
+					}
+				}
+
+				// Clean up toggles for non-existent files
+				for (const togglePath in updatedToggles) {
+					const pathExists = existingRulePaths.has(togglePath)
+					if (!pathExists) {
+						delete updatedToggles[togglePath]
+					}
+				}
+			} else {
+				// FILE CASE
+				// Add toggle for this file
+				const pathHasToggle = rulesDirectoryPath in updatedToggles
+				if (!pathHasToggle) {
+					updatedToggles[rulesDirectoryPath] = true
+				}
+
+				// Remove toggles for any other paths
+				for (const togglePath in updatedToggles) {
+					if (togglePath !== rulesDirectoryPath) {
+						delete updatedToggles[togglePath]
+					}
+				}
+			}
+		} else {
+			// PATH DOESN'T EXIST CASE
+			// Clear all toggles since the path doesn't exist
+			for (const togglePath in updatedToggles) {
+				delete updatedToggles[togglePath]
+			}
+		}
+	} catch (error) {
+		console.error(`Failed to synchronize rule toggles for path: ${rulesDirectoryPath}`, error)
+	}
+
+	return updatedToggles
+}
+
+export async function refreshClineRulesToggles(
+	context: vscode.ExtensionContext,
+	workingDirectory: string,
+): Promise<{
+	globalToggles: ClineRulesToggles
+	localToggles: ClineRulesToggles
+}> {
+	// Global toggles
+	const globalClineRulesToggles = ((await getGlobalState(context, "globalClineRulesToggles")) as ClineRulesToggles) || {}
+	const globalClineRulesFilePath = await ensureRulesDirectoryExists()
+	const updatedGlobalToggles = await synchronizeRuleToggles(globalClineRulesFilePath, globalClineRulesToggles)
+	await updateGlobalState(context, "globalClineRulesToggles", updatedGlobalToggles)
+
+	// Local toggles
+	const localClineRulesToggles = ((await getWorkspaceState(context, "localClineRulesToggles")) as ClineRulesToggles) || {}
+	const localClineRulesFilePath = path.resolve(workingDirectory, GlobalFileNames.clineRules)
+	const updatedLocalToggles = await synchronizeRuleToggles(localClineRulesFilePath, localClineRulesToggles)
+	await updateWorkspaceState(context, "localClineRulesToggles", updatedLocalToggles)
+
+	return {
+		globalToggles: updatedGlobalToggles,
+		localToggles: updatedLocalToggles,
+	}
+}

src/core/controller/browser/getBrowserConnectionInfo.ts

@@ -0,0 +1,47 @@
+import { BrowserConnectionInfo } from "../../../shared/proto/browser"
+import { EmptyRequest } from "../../../shared/proto/common"
+import { Controller } from "../index"
+import { getAllExtensionState } from "../../storage/state"
+
+/**
+ * Get information about the current browser connection
+ * @param controller The controller instance
+ * @param request The request message
+ * @returns The browser connection info
+ */
+export async function getBrowserConnectionInfo(controller: Controller, request: EmptyRequest): Promise<BrowserConnectionInfo> {
+	try {
+		// Get browser settings from extension state
+		const { browserSettings } = await getAllExtensionState(controller.context)
+
+		// Check if there's an active browser session by using the controller's handleWebviewMessage approach
+		// This is similar to what's done in controller/index.ts for the "getBrowserConnectionInfo" message
+		if (controller.task?.browserSession) {
+			// Access the browser session through the controller's task property
+			// Using indexer notation to access private property
+			const browserSession = controller.task.browserSession
+			const connectionInfo = browserSession.getConnectionInfo()
+
+			// Convert from BrowserSession.BrowserConnectionInfo to proto.BrowserConnectionInfo
+			return {
+				isConnected: connectionInfo.isConnected,
+				isRemote: connectionInfo.isRemote,
+				host: connectionInfo.host || "", // Ensure host is never undefined
+			}
+		}
+
+		// Fallback to browser settings if no active browser session
+		return {
+			isConnected: false,
+			isRemote: !!browserSettings.remoteBrowserEnabled,
+			host: browserSettings.remoteBrowserHost || "",
+		}
+	} catch (error: unknown) {
+		console.error("Error getting browser connection info:", error)
+		return {
+			isConnected: false,
+			isRemote: false,
+			host: "",
+		}
+	}
+}

src/core/controller/browser/index.ts

@@ -0,0 +1,15 @@
+import { createServiceRegistry, ServiceMethodHandler } from "../grpc-service"
+import { registerAllMethods } from "./methods"
+
+// Create browser service registry
+const browserService = createServiceRegistry("browser")
+
+// Export the method handler type and registration function
+export type BrowserMethodHandler = ServiceMethodHandler
+export const registerMethod = browserService.registerMethod
+
+// Export the request handler
+export const handleBrowserServiceRequest = browserService.handleRequest
+
+// Register all browser methods
+registerAllMethods()

src/core/controller/browser/methods.ts

@@ -0,0 +1,14 @@
+// AUTO-GENERATED FILE - DO NOT MODIFY DIRECTLY
+// Generated by proto/build-proto.js
+
+// Import all method implementations
+import { registerMethod } from "./index"
+import { getBrowserConnectionInfo } from "./getBrowserConnectionInfo"
+import { testBrowserConnection } from "./testBrowserConnection"
+
+// Register all browser service methods
+export function registerAllMethods(): void {
+	// Register each method with the registry
+	registerMethod("getBrowserConnectionInfo", getBrowserConnectionInfo)
+	registerMethod("testBrowserConnection", testBrowserConnection)
+}

src/core/controller/browser/testBrowserConnection.ts

@@ -0,0 +1,63 @@
+import { BrowserConnection } from "../../../shared/proto/browser"
+import { StringRequest } from "../../../shared/proto/common"
+import { Controller } from "../index"
+import { getAllExtensionState } from "../../storage/state"
+import { BrowserSession } from "../../../services/browser/BrowserSession"
+import { discoverChromeInstances } from "../../../services/browser/BrowserDiscovery"
+
+/**
+ * Test connection to a browser instance
+ * @param controller The controller instance
+ * @param request The request message
+ * @returns The browser connection result
+ */
+export async function testBrowserConnection(controller: Controller, request: StringRequest): Promise<BrowserConnection> {
+	try {
+		const { browserSettings } = await getAllExtensionState(controller.context)
+		const browserSession = new BrowserSession(controller.context, browserSettings)
+		const text = request.value || ""
+
+		// If no text is provided, try auto-discovery
+		if (!text) {
+			try {
+				const discoveredHost = await discoverChromeInstances()
+				if (discoveredHost) {
+					// Test the connection to the discovered host
+					const result = await browserSession.testConnection(discoveredHost)
+					return {
+						success: result.success,
+						message: `Auto-discovered and tested connection to Chrome at ${discoveredHost}: ${result.message}`,
+						endpoint: result.endpoint || "",
+					}
+				} else {
+					return {
+						success: false,
+						message:
+							"No Chrome instances found on the network. Make sure Chrome is running with remote debugging enabled (--remote-debugging-port=9222).",
+						endpoint: "",
+					}
+				}
+			} catch (error) {
+				return {
+					success: false,
+					message: `Error during auto-discovery: ${error instanceof Error ? error.message : String(error)}`,
+					endpoint: "",
+				}
+			}
+		} else {
+			// Test the provided URL
+			const result = await browserSession.testConnection(text)
+			return {
+				success: result.success,
+				message: result.message,
+				endpoint: result.endpoint || "",
+			}
+		}
+	} catch (error) {
+		return {
+			success: false,
+			message: `Error testing connection: ${error instanceof Error ? error.message : String(error)}`,
+			endpoint: "",
+		}
+	}
+}

src/core/controller/grpc-handler.ts

@@ -0,0 +1,81 @@
+import { Controller } from "./index"
+import { handleBrowserServiceRequest } from "./browser/index"
+import { ExtensionMessage } from "../../shared/ExtensionMessage"
+
+/**
+ * Handles gRPC requests from the webview
+ */
+export class GrpcHandler {
+	constructor(private controller: Controller) {}
+
+	/**
+	 * Handle a gRPC request from the webview
+	 * @param service The service name
+	 * @param method The method name
+	 * @param message The request message
+	 * @param requestId The request ID for response correlation
+	 * @returns The response message or error
+	 */
+	async handleRequest(
+		service: string,
+		method: string,
+		message: any,
+		requestId: string,
+	): Promise<{
+		message?: any
+		error?: string
+		request_id: string
+	}> {
+		try {
+			// Handle BrowserService requests
+			if (service === "cline.BrowserService") {
+				return {
+					message: await handleBrowserServiceRequest(this.controller, method, message),
+					request_id: requestId,
+				}
+			}
+
+			throw new Error(`Unknown service: ${service}`)
+		} catch (error) {
+			return {
+				error: error instanceof Error ? error.message : String(error),
+				request_id: requestId,
+			}
+		}
+	}
+}
+
+/**
+ * Handle a gRPC request from the webview
+ * @param controller The controller instance
+ * @param request The gRPC request
+ */
+export async function handleGrpcRequest(
+	controller: Controller,
+	request: {
+		service: string
+		method: string
+		message: any
+		request_id: string
+	},
+) {
+	try {
+		const grpcHandler = new GrpcHandler(controller)
+		const response = await grpcHandler.handleRequest(request.service, request.method, request.message, request.request_id)
+
+		// Send the response back to the webview
+		await controller.postMessageToWebview({
+			type: "grpc_response",
+			grpc_response: response,
+		})
+	} catch (error) {
+		// Send error response
+		await controller.postMessageToWebview({
+			type: "grpc_response",
+			grpc_response: {
+				error: error instanceof Error ? error.message : String(error),
+				request_id: request.request_id,
+			},
+		})
+	}
+}

src/core/controller/grpc-service.ts

@@ -0,0 +1,65 @@
+import { Controller } from "./index"
+
+/**
+ * Generic type for service method handlers
+ */
+export type ServiceMethodHandler = (controller: Controller, message: any) => Promise<any>
+
+/**
+ * Generic service registry for gRPC services
+ */
+export class ServiceRegistry {
+	private serviceName: string
+	private methodRegistry: Record<string, ServiceMethodHandler> = {}
+
+	/**
+	 * Create a new service registry
+	 * @param serviceName The name of the service (used for logging)
+	 */
+	constructor(serviceName: string) {
+		this.serviceName = serviceName
+	}
+
+	/**
+	 * Register a method handler
+	 * @param methodName The name of the method to register
+	 * @param handler The handler function for the method
+	 */
+	registerMethod(methodName: string, handler: ServiceMethodHandler): void {
+		this.methodRegistry[methodName] = handler
+		console.log(`Registered ${this.serviceName} method: ${methodName}`)
+	}
+
+	/**
+	 * Handle a service request
+	 * @param controller The controller instance
+	 * @param method The method name
+	 * @param message The request message
+	 * @returns The response message
+	 */
+	async handleRequest(controller: Controller, method: string, message: any): Promise<any> {
+		const handler = this.methodRegistry[method]
+
+		if (!handler) {
+			throw new Error(`Unknown ${this.serviceName} method: ${method}`)
+		}
+
+		return handler(controller, message)
+	}
+}
+
+/**
+ * Create a service registry factory function
+ * @param serviceName The name of the service
+ * @returns An object with register and handle functions
+ */
+export function createServiceRegistry(serviceName: string) {
+	const registry = new ServiceRegistry(serviceName)
+
+	return {
+		registerMethod: (methodName: string, handler: ServiceMethodHandler) => registry.registerMethod(methodName, handler),
+
+		handleRequest: (controller: Controller, method: string, message: any) =>
+			registry.handleRequest(controller, method, message),
+	}
+}

src/core/controller/index.ts

@@ -2,13 +2,12 @@ import { Anthropic } from "@anthropic-ai/sdk"
 import axios from "axios"
 import type { AxiosRequestConfig } from "axios"
 import crypto from "crypto"
-import { execa } from "execa"
 import fs from "fs/promises"
 import { setTimeout as setTimeoutPromise } from "node:timers/promises"
-import os from "os"
 import pWaitFor from "p-wait-for"
 import * as path from "path"
 import * as vscode from "vscode"
+import { handleGrpcRequest } from "./grpc-handler"
 import { buildApiHandler } from "../../api"
 import { cleanupLegacyCheckpoints } from "../../integrations/checkpoints/CheckpointMigration"
 import { downloadTask } from "../../integrations/misc/export-markdown"
@@ -36,17 +35,21 @@ import { searchCommits } from "../../utils/git"
 import { getWorkspacePath } from "../../utils/path"
 import { getTotalTasksSize } from "../../utils/storage"
 import { openMention } from "../mentions"
-import { GlobalFileNames } from "../storage/disk"
+import { ensureMcpServersDirectoryExists, ensureSettingsDirectoryExists, GlobalFileNames } from "../storage/disk"
 import {
 	getAllExtensionState,
 	getGlobalState,
 	getSecret,
+	getWorkspaceState,
 	resetExtensionState,
 	storeSecret,
 	updateApiConfiguration,
 	updateGlobalState,
+	updateWorkspaceState,
 } from "../storage/state"
-import { Task } from "../task"
+import { Task, cwd } from "../task"
+import { ClineRulesToggles } from "../../shared/cline-rules"
+import { refreshClineRulesToggles } from "../context/instructions/user-instructions/cline-rules"
 
 /*
 https://github.com/microsoft/vscode-webview-ui-toolkit-samples/blob/main/default/weather-webview/src/providers/WeatherViewProvider.ts
@@ -55,27 +58,27 @@ https://github.com/KumarVariable/vscode-extension-sidebar-html/blob/master/src/c
 */
 
 export class Controller {
-	private postMessage: (message: any) => Thenable<boolean> | undefined
+	private postMessage: (message: ExtensionMessage) => Thenable<boolean> | undefined
 
 	private disposables: vscode.Disposable[] = []
-	private task?: Task
+	task?: Task
 	workspaceTracker: WorkspaceTracker
 	mcpHub: McpHub
 	accountService: ClineAccountService
-	private latestAnnouncementId = "april-10-2025" // update to some unique identifier when we add a new announcement
+	private latestAnnouncementId = "april-11-2025" // update to some unique identifier when we add a new announcement
 
 	constructor(
 		readonly context: vscode.ExtensionContext,
 		private readonly outputChannel: vscode.OutputChannel,
-		postMessage: (message: any) => Thenable<boolean> | undefined,
+		postMessage: (message: ExtensionMessage) => Thenable<boolean> | undefined,
 	) {
 		this.outputChannel.appendLine("ClineProvider instantiated")
 		this.postMessage = postMessage
 
 		this.workspaceTracker = new WorkspaceTracker((msg) => this.postMessageToWebview(msg))
 		this.mcpHub = new McpHub(
-			() => this.ensureMcpServersDirectoryExists(),
-			() => this.ensureSettingsDirectoryExists(),
+			() => ensureMcpServersDirectoryExists(),
+			() => ensureSettingsDirectoryExists(this.context),
 			(msg) => this.postMessageToWebview(msg),
 			this.context.extension?.packageJSON?.version ?? "1.0.0",
 		)
@@ -256,6 +259,13 @@ export class Controller {
 					telemetryService.updateTelemetryState(isOptedIn)
 				})
 				break
+			case "showChatView": {
+				this.postMessageToWebview({
+					type: "action",
+					action: "chatButtonClicked",
+				})
+				break
+			}
 			case "newTask":
 				// Code that should run in response to the hello message command
 				//vscode.window.showInformationMessage(message.text!)
@@ -301,88 +311,6 @@ export class Controller {
 					await this.postStateToWebview()
 				}
 				break
-			case "getBrowserConnectionInfo":
-				try {
-					// Get the current browser session from Cline if it exists
-					if (this.task?.browserSession) {
-						const connectionInfo = this.task.browserSession.getConnectionInfo()
-						await this.postMessageToWebview({
-							type: "browserConnectionInfo",
-							isConnected: connectionInfo.isConnected,
-							isRemote: connectionInfo.isRemote,
-							host: connectionInfo.host,
-						})
-					} else {
-						// If no active browser session, just return the settings
-						const { browserSettings } = await getAllExtensionState(this.context)
-						await this.postMessageToWebview({
-							type: "browserConnectionInfo",
-							isConnected: false,
-							isRemote: !!browserSettings.remoteBrowserEnabled,
-							host: browserSettings.remoteBrowserHost,
-						})
-					}
-				} catch (error) {
-					console.error("Error getting browser connection info:", error)
-					await this.postMessageToWebview({
-						type: "browserConnectionInfo",
-						isConnected: false,
-						isRemote: false,
-					})
-				}
-				break
-			case "testBrowserConnection":
-				try {
-					const { browserSettings } = await getAllExtensionState(this.context)
-					const browserSession = new BrowserSession(this.context, browserSettings)
-					// If no text is provided, try auto-discovery
-					if (!message.text) {
-						try {
-							const discoveredHost = await discoverChromeInstances()
-							if (discoveredHost) {
-								// Test the connection to the discovered host
-								const result = await browserSession.testConnection(discoveredHost)
-								// Send the result back to the webview
-								await this.postMessageToWebview({
-									type: "browserConnectionResult",
-									success: result.success,
-									text: `Auto-discovered and tested connection to Chrome at ${discoveredHost}: ${result.message}`,
-									endpoint: result.endpoint,
-								})
-							} else {
-								await this.postMessageToWebview({
-									type: "browserConnectionResult",
-									success: false,
-									text: "No Chrome instances found on the network. Make sure Chrome is running with remote debugging enabled (--remote-debugging-port=9222).",
-								})
-							}
-						} catch (error) {
-							await this.postMessageToWebview({
-								type: "browserConnectionResult",
-								success: false,
-								text: `Error during auto-discovery: ${error instanceof Error ? error.message : String(error)}`,
-							})
-						}
-					} else {
-						// Test the provided URL
-						const result = await browserSession.testConnection(message.text)
-
-						// Send the result back to the webview
-						await this.postMessageToWebview({
-							type: "browserConnectionResult",
-							success: result.success,
-							text: result.message,
-							endpoint: result.endpoint,
-						})
-					}
-				} catch (error) {
-					await this.postMessageToWebview({
-						type: "browserConnectionResult",
-						success: false,
-						text: `Error testing connection: ${error instanceof Error ? error.message : String(error)}`,
-					})
-				}
-				break
 			case "discoverBrowser":
 				try {
 					const discoveredHost = await discoverChromeInstances()
@@ -501,6 +429,10 @@ export class Controller {
 				const openAiModels = await this.getOpenAiModels(apiConfiguration.openAiBaseUrl, apiConfiguration.openAiApiKey)
 				this.postMessageToWebview({ type: "openAiModels", openAiModels })
 				break
+			case "refreshClineRules":
+				await refreshClineRulesToggles(this.context, cwd)
+				await this.postStateToWebview()
+				break
 			case "openImage":
 				openImage(message.text!)
 				break
@@ -528,7 +460,7 @@ export class Controller {
 				break
 			}
 			case "checkpointRestore": {
-				await this.cancelTask() // we cannot alter message history say if the task is active, as it could be in the middle of editing a file or running a command, which expect the ask to be responded to rather than being superceded by a new message eg add deleted_api_reqs
+				await this.cancelTask() // we cannot alter message history say if the task is active, as it could be in the middle of editing a file or running a command, which expect the ask to be responded to rather than being superseded by a new message eg add deleted_api_reqs
 				// cancel task waits for any open editor to be reverted and starts a new cline instance
 				if (message.number) {
 					// wait for messages to be loaded
@@ -538,7 +470,7 @@ export class Controller {
 						console.error("Failed to init new cline instance")
 					})
 					// NOTE: cancelTask awaits abortTask, which awaits diffViewProvider.revertChanges, which reverts any edited files, allowing us to reset to a checkpoint rather than running into a state where the revertChanges function is called alongside or after the checkpoint reset
-					await this.task?.restoreCheckpoint(message.number, message.text! as ClineCheckpointRestore)
+					await this.task?.restoreCheckpoint(message.number, message.text! as ClineCheckpointRestore, message.offset)
 				}
 				break
 			}
@@ -606,14 +538,7 @@ export class Controller {
 						await this.togglePlanActModeWithChatSettings({ mode: "act" })
 					}
 
-					// 2. Enable MCP settings if disabled
-					// Enable MCP mode if disabled
-					const mcpConfig = vscode.workspace.getConfiguration("cline.mcp")
-					if (mcpConfig.get<string>("mode") !== "full") {
-						await mcpConfig.update("mode", "full", true)
-					}
-
-					// 3. download MCP
+					// 2. download MCP
 					await this.downloadMcp(message.mcpId)
 				}
 				break
@@ -685,6 +610,30 @@ export class Controller {
 				}
 				break
 			}
+			case "toggleClineRule": {
+				const { isGlobal, rulePath, enabled } = message
+				if (rulePath && typeof enabled === "boolean" && typeof isGlobal === "boolean") {
+					if (isGlobal) {
+						const toggles =
+							((await getGlobalState(this.context, "globalClineRulesToggles")) as ClineRulesToggles) || {}
+						toggles[rulePath] = enabled
+						await updateGlobalState(this.context, "globalClineRulesToggles", toggles)
+					} else {
+						const toggles =
+							((await getWorkspaceState(this.context, "localClineRulesToggles")) as ClineRulesToggles) || {}
+						toggles[rulePath] = enabled
+						await updateWorkspaceState(this.context, "localClineRulesToggles", toggles)
+					}
+					await this.postStateToWebview()
+				} else {
+					console.error("toggleClineRule: Missing or invalid parameters", {
+						rulePath,
+						isGlobal: typeof isGlobal === "boolean" ? isGlobal : `Invalid: ${typeof isGlobal}`,
+						enabled: typeof enabled === "boolean" ? enabled : `Invalid: ${typeof enabled}`,
+					})
+				}
+				break
+			}
 			case "requestTotalTasksSize": {
 				this.refreshTotalTasksSize()
 				break
@@ -919,6 +868,12 @@ export class Controller {
 				}
 				break
 			}
+			case "grpc_request": {
+				if (message.grpc_request) {
+					await handleGrpcRequest(this, message.grpc_request)
+				}
+				break
+			}
 			// Add more switch case statements here as more webview message commands
 			// are created within the webview context (i.e. inside media/main.js)
 		}
@@ -1113,61 +1068,6 @@ export class Controller {
 		}
 	}
 
-	// MCP
-
-	async getDocumentsPath(): Promise<string> {
-		if (process.platform === "win32") {
-			try {
-				const { stdout: docsPath } = await execa("powershell", [
-					"-NoProfile", // Ignore user's PowerShell profile(s)
-					"-Command",
-					"[System.Environment]::GetFolderPath([System.Environment+SpecialFolder]::MyDocuments)",
-				])
-				const trimmedPath = docsPath.trim()
-				if (trimmedPath) {
-					return trimmedPath
-				}
-			} catch (err) {
-				console.error("Failed to retrieve Windows Documents path. Falling back to homedir/Documents.")
-			}
-		} else if (process.platform === "linux") {
-			try {
-				// First check if xdg-user-dir exists
-				await execa("which", ["xdg-user-dir"])
-
-				// If it exists, try to get XDG documents path
-				const { stdout } = await execa("xdg-user-dir", ["DOCUMENTS"])
-				const trimmedPath = stdout.trim()
-				if (trimmedPath) {
-					return trimmedPath
-				}
-			} catch {
-				// Log error but continue to fallback
-				console.error("Failed to retrieve XDG Documents path. Falling back to homedir/Documents.")
-			}
-		}
-
-		// Default fallback for all platforms
-		return path.join(os.homedir(), "Documents")
-	}
-
-	async ensureMcpServersDirectoryExists(): Promise<string> {
-		const userDocumentsPath = await this.getDocumentsPath()
-		const mcpServersDir = path.join(userDocumentsPath, "Cline", "MCP")
-		try {
-			await fs.mkdir(mcpServersDir, { recursive: true })
-		} catch (error) {
-			return "~/Documents/Cline/MCP" // in case creating a directory in documents fails for whatever reason (e.g. permissions) - this is fine since this path is only ever used in the system prompt
-		}
-		return mcpServersDir
-	}
-
-	async ensureSettingsDirectoryExists(): Promise<string> {
-		const settingsDir = path.join(this.context.globalStorageUri.fsPath, "settings")
-		await fs.mkdir(settingsDir, { recursive: true })
-		return settingsDir
-	}
-
 	// VSCode LM API
 
 	private async getVsCodeLmModels() {
@@ -1409,6 +1309,7 @@ export class Controller {
 
 			// Create task with context from README and added guidelines for MCP server installation
 			const task = `Set up the MCP server from ${mcpDetails.githubUrl} while adhering to these MCP server installation rules:
+- Start by loading the MCP documentation.
 - Use "${mcpDetails.mcpId}" as the server name in cline_mcp_settings.json.
 - Create the directory for the new MCP server before starting installation.
 - Make sure you read the user's existing cline_mcp_settings.json file before editing it with this new mcp, to not overwrite any existing servers.
@@ -1819,7 +1720,7 @@ Here is the project's README to help you get started:\n\n${mcpDetails.readmeCont
 			}
 		}
 		// if we tried to get a task that doesn't exist, remove it from state
-		// FIXME: this seems to happen sometimes when the json file doesnt save to disk for some reason
+		// FIXME: this seems to happen sometimes when the json file doesn't save to disk for some reason
 		await this.deleteTaskFromState(id)
 		throw new Error("Task not found")
 	}
@@ -1951,8 +1852,12 @@ Here is the project's README to help you get started:\n\n${mcpDetails.readmeCont
 			mcpMarketplaceEnabled,
 			telemetrySetting,
 			planActSeparateModelsSetting,
+			globalClineRulesToggles,
 		} = await getAllExtensionState(this.context)
 
+		const localClineRulesToggles =
+			((await getWorkspaceState(this.context, "localClineRulesToggles")) as ClineRulesToggles) || {}
+
 		return {
 			version: this.context.extension?.packageJSON?.version ?? "",
 			apiConfiguration,
@@ -1975,6 +1880,8 @@ Here is the project's README to help you get started:\n\n${mcpDetails.readmeCont
 			telemetrySetting,
 			planActSeparateModelsSetting,
 			vscMachineId: vscode.env.machineId,
+			globalClineRulesToggles: globalClineRulesToggles || {},
+			localClineRulesToggles: localClineRulesToggles || {},
 		}
 	}
 
@@ -1996,7 +1903,7 @@ Here is the project's README to help you get started:\n\n${mcpDetails.readmeCont
 	// conversation history to send in API requests
 
 	/*
-	It seems that some API messages do not comply with vscode state requirements. Either the Anthropic library is manipulating these values somehow in the backend in a way thats creating cyclic references, or the API returns a function or a Symbol as part of the message content.
+	It seems that some API messages do not comply with vscode state requirements. Either the Anthropic library is manipulating these values somehow in the backend in a way that's creating cyclic references, or the API returns a function or a Symbol as part of the message content.
 	VSCode docs about state: "The value must be JSON-stringifyable ... value — A value. MUST not contain cyclic references."
 	For now we'll store the conversation history in memory, and if we need to store in state directly we'd need to do a manual conversion to ensure proper json stringification.
 	*/

src/core/prompts/loadMcpDocumentation.ts

@@ -0,0 +1,361 @@
+import { McpHub } from "../../services/mcp/McpHub"
+
+export async function loadMcpDocumentation(mcpHub: McpHub) {
+	return `## Creating an MCP Server
+
+When creating MCP servers, it's important to understand that they operate in a non-interactive environment. The server cannot initiate OAuth flows, open browser windows, or prompt for user input during runtime. All credentials and authentication tokens must be provided upfront through environment variables in the MCP settings configuration. For example, Spotify's API uses OAuth to get a refresh token for the user, but the MCP server cannot initiate this flow. While you can walk the user through obtaining an application client ID and secret, you may have to create a separate one-time setup script (like get-refresh-token.js) that captures and logs the final piece of the puzzle: the user's refresh token (i.e. you might run the script using execute_command which would open a browser for authentication, and then log the refresh token so that you can see it in the command output for you to use in the MCP settings configuration).
+
+Unless the user specifies otherwise, new MCP servers should be created in: ${await mcpHub.getMcpServersPath()}
+
+### Example MCP Server
+
+For example, if the user wanted to give you the ability to retrieve weather information, you could create an MCP server that uses the OpenWeather API to get weather information, add it to the MCP settings configuration file, and then notice that you now have access to new tools and resources in the system prompt that you might use to show the user your new capabilities.
+
+The following example demonstrates how to build an MCP server that provides weather data functionality. While this example shows how to implement resources, resource templates, and tools, in practice you should prefer using tools since they are more flexible and can handle dynamic parameters. The resource and resource template implementations are included here mainly for demonstration purposes of the different MCP capabilities, but a real weather server would likely just expose tools for fetching weather data. (The following steps are for macOS)
+
+1. Use the \`create-typescript-server\` tool to bootstrap a new project in the default MCP servers directory:
+
+\`\`\`bash
+cd ${await mcpHub.getMcpServersPath()}
+npx @modelcontextprotocol/create-server weather-server
+cd weather-server
+# Install dependencies
+npm install axios
+\`\`\`
+
+This will create a new project with the following structure:
+
+\`\`\`
+weather-server/
+  ├── package.json
+      {
+        ...
+        "type": "module", // added by default, uses ES module syntax (import/export) rather than CommonJS (require/module.exports) (Important to know if you create additional scripts in this server repository like a get-refresh-token.js script)
+        "scripts": {
+          "build": "tsc && node -e \"require('fs').chmodSync('build/index.js', '755')\"",
+          ...
+        }
+        ...
+      }
+  ├── tsconfig.json
+  └── src/
+      └── weather-server/
+          └── index.ts      # Main server implementation
+\`\`\`
+
+2. Replace \`src/index.ts\` with the following:
+
+\`\`\`typescript
+#!/usr/bin/env node
+import { Server } from '@modelcontextprotocol/sdk/server/index.js';
+import { StdioServerTransport } from '@modelcontextprotocol/sdk/server/stdio.js';
+import {
+  CallToolRequestSchema,
+  ErrorCode,
+  ListResourcesRequestSchema,
+  ListResourceTemplatesRequestSchema,
+  ListToolsRequestSchema,
+  McpError,
+  ReadResourceRequestSchema,
+} from '@modelcontextprotocol/sdk/types.js';
+import axios from 'axios';
+
+const API_KEY = process.env.OPENWEATHER_API_KEY; // provided by MCP config
+if (!API_KEY) {
+  throw new Error('OPENWEATHER_API_KEY environment variable is required');
+}
+
+interface OpenWeatherResponse {
+  main: {
+    temp: number;
+    humidity: number;
+  };
+  weather: [{ description: string }];
+  wind: { speed: number };
+  dt_txt?: string;
+}
+
+const isValidForecastArgs = (
+  args: any
+): args is { city: string; days?: number } =>
+  typeof args === 'object' &&
+  args !== null &&
+  typeof args.city === 'string' &&
+  (args.days === undefined || typeof args.days === 'number');
+
+class WeatherServer {
+  private server: Server;
+  private axiosInstance;
+
+  constructor() {
+    this.server = new Server(
+      {
+        name: 'example-weather-server',
+        version: '0.1.0',
+      },
+      {
+        capabilities: {
+          resources: {},
+          tools: {},
+        },
+      }
+    );
+
+    this.axiosInstance = axios.create({
+      baseURL: 'http://api.openweathermap.org/data/2.5',
+      params: {
+        appid: API_KEY,
+        units: 'metric',
+      },
+    });
+
+    this.setupResourceHandlers();
+    this.setupToolHandlers();
+    
+    // Error handling
+    this.server.onerror = (error) => console.error('[MCP Error]', error);
+    process.on('SIGINT', async () => {
+      await this.server.close();
+      process.exit(0);
+    });
+  }
+
+  // MCP Resources represent any kind of UTF-8 encoded data that an MCP server wants to make available to clients, such as database records, API responses, log files, and more. Servers define direct resources with a static URI or dynamic resources with a URI template that follows the format \`[protocol]://[host]/[path]\`.
+  private setupResourceHandlers() {
+    // For static resources, servers can expose a list of resources:
+    this.server.setRequestHandler(ListResourcesRequestSchema, async () => ({
+      resources: [
+        // This is a poor example since you could use the resource template to get the same information but this demonstrates how to define a static resource
+        {
+          uri: \`weather://San Francisco/current\`, // Unique identifier for San Francisco weather resource
+          name: \`Current weather in San Francisco\`, // Human-readable name
+          mimeType: 'application/json', // Optional MIME type
+          // Optional description
+          description:
+            'Real-time weather data for San Francisco including temperature, conditions, humidity, and wind speed',
+        },
+      ],
+    }));
+
+    // For dynamic resources, servers can expose resource templates:
+    this.server.setRequestHandler(
+      ListResourceTemplatesRequestSchema,
+      async () => ({
+        resourceTemplates: [
+          {
+            uriTemplate: 'weather://{city}/current', // URI template (RFC 6570)
+            name: 'Current weather for a given city', // Human-readable name
+            mimeType: 'application/json', // Optional MIME type
+            description: 'Real-time weather data for a specified city', // Optional description
+          },
+        ],
+      })
+    );
+
+    // ReadResourceRequestSchema is used for both static resources and dynamic resource templates
+    this.server.setRequestHandler(
+      ReadResourceRequestSchema,
+      async (request) => {
+        const match = request.params.uri.match(
+          /^weather:\/\/([^/]+)\/current$/
+        );
+        if (!match) {
+          throw new McpError(
+            ErrorCode.InvalidRequest,
+            \`Invalid URI format: \${request.params.uri}\`
+          );
+        }
+        const city = decodeURIComponent(match[1]);
+
+        try {
+          const response = await this.axiosInstance.get(
+            'weather', // current weather
+            {
+              params: { q: city },
+            }
+          );
+
+          return {
+            contents: [
+              {
+                uri: request.params.uri,
+                mimeType: 'application/json',
+                text: JSON.stringify(
+                  {
+                    temperature: response.data.main.temp,
+                    conditions: response.data.weather[0].description,
+                    humidity: response.data.main.humidity,
+                    wind_speed: response.data.wind.speed,
+                    timestamp: new Date().toISOString(),
+                  },
+                  null,
+                  2
+                ),
+              },
+            ],
+          };
+        } catch (error) {
+          if (axios.isAxiosError(error)) {
+            throw new McpError(
+              ErrorCode.InternalError,
+              \`Weather API error: \${
+                error.response?.data.message ?? error.message
+              }\`
+            );
+          }
+          throw error;
+        }
+      }
+    );
+  }
+
+  /* MCP Tools enable servers to expose executable functionality to the system. Through these tools, you can interact with external systems, perform computations, and take actions in the real world.
+   * - Like resources, tools are identified by unique names and can include descriptions to guide their usage. However, unlike resources, tools represent dynamic operations that can modify state or interact with external systems.
+   * - While resources and tools are similar, you should prefer to create tools over resources when possible as they provide more flexibility.
+   */
+  private setupToolHandlers() {
+    this.server.setRequestHandler(ListToolsRequestSchema, async () => ({
+      tools: [
+        {
+          name: 'get_forecast', // Unique identifier
+          description: 'Get weather forecast for a city', // Human-readable description
+          inputSchema: {
+            // JSON Schema for parameters
+            type: 'object',
+            properties: {
+              city: {
+                type: 'string',
+                description: 'City name',
+              },
+              days: {
+                type: 'number',
+                description: 'Number of days (1-5)',
+                minimum: 1,
+                maximum: 5,
+              },
+            },
+            required: ['city'], // Array of required property names
+          },
+        },
+      ],
+    }));
+
+    this.server.setRequestHandler(CallToolRequestSchema, async (request) => {
+      if (request.params.name !== 'get_forecast') {
+        throw new McpError(
+          ErrorCode.MethodNotFound,
+          \`Unknown tool: \${request.params.name}\`
+        );
+      }
+
+      if (!isValidForecastArgs(request.params.arguments)) {
+        throw new McpError(
+          ErrorCode.InvalidParams,
+          'Invalid forecast arguments'
+        );
+      }
+
+      const city = request.params.arguments.city;
+      const days = Math.min(request.params.arguments.days || 3, 5);
+
+      try {
+        const response = await this.axiosInstance.get<{
+          list: OpenWeatherResponse[];
+        }>('forecast', {
+          params: {
+            q: city,
+            cnt: days * 8,
+          },
+        });
+
+        return {
+          content: [
+            {
+              type: 'text',
+              text: JSON.stringify(response.data.list, null, 2),
+            },
+          ],
+        };
+      } catch (error) {
+        if (axios.isAxiosError(error)) {
+          return {
+            content: [
+              {
+                type: 'text',
+                text: \`Weather API error: \${
+                  error.response?.data.message ?? error.message
+                }\`,
+              },
+            ],
+            isError: true,
+          };
+        }
+        throw error;
+      }
+    });
+  }
+
+  async run() {
+    const transport = new StdioServerTransport();
+    await this.server.connect(transport);
+    console.error('Weather MCP server running on stdio');
+  }
+}
+
+const server = new WeatherServer();
+server.run().catch(console.error);
+\`\`\`
+
+(Remember: This is just an example–you may use different dependencies, break the implementation up into multiple files, etc.)
+
+3. Build and compile the executable JavaScript file
+
+\`\`\`bash
+npm run build
+\`\`\`
+
+4. Whenever you need an environment variable such as an API key to configure the MCP server, walk the user through the process of getting the key. For example, they may need to create an account and go to a developer dashboard to generate the key. Provide step-by-step instructions and URLs to make it easy for the user to retrieve the necessary information. Then use the ask_followup_question tool to ask the user for the key, in this case the OpenWeather API key.
+
+5. Install the MCP Server by adding the MCP server configuration to the settings file located at '${await mcpHub.getMcpSettingsFilePath()}'. The settings file may have other MCP servers already configured, so you would read it first and then add your new server to the existing \`mcpServers\` object.
+
+IMPORTANT: Regardless of what else you see in the MCP settings file, you must default any new MCP servers you create to disabled=false and autoApprove=[].
+
+\`\`\`json
+{
+  "mcpServers": {
+    ...,
+    "weather": {
+      "command": "node",
+      "args": ["/path/to/weather-server/build/index.js"],
+      "env": {
+        "OPENWEATHER_API_KEY": "user-provided-api-key"
+      }
+    },
+  }
+}
+\`\`\`
+
+(Note: the user may also ask you to install the MCP server to the Claude desktop app, in which case you would read then modify \`~/Library/Application\ Support/Claude/claude_desktop_config.json\` on macOS for example. It follows the same format of a top level \`mcpServers\` object.)
+
+6. After you have edited the MCP settings configuration file, the system will automatically run all the servers and expose the available tools and resources in the 'Connected MCP Servers' section. (Note: If you encounter a 'not connected' error when testing a newly installed mcp server, a common cause is an incorrect build path in your MCP settings configuration. Since compiled JavaScript files are commonly output to either 'dist/' or 'build/' directories, double-check that the build path in your MCP settings matches where your files are actually being compiled. E.g. If you assumed 'build' as the folder, check tsconfig.json to see if it's using 'dist' instead.)
+
+7. Now that you have access to these new tools and resources, you may suggest ways the user can command you to invoke them - for example, with this new weather tool now available, you can invite the user to ask "what's the weather in San Francisco?"
+
+## Editing MCP Servers
+
+The user may ask to add tools or resources that may make sense to add to an existing MCP server (listed under 'Connected MCP Servers' below: ${
+		mcpHub
+			.getServers()
+			.filter((server) => server.status === "connected")
+			.map((server) => server.name)
+			.join(", ") || "(None running currently)"
+	}, e.g. if it would use the same API. This would be possible if you can locate the MCP server repository on the user's system by looking at the server arguments for a filepath. You might then use list_files and read_file to explore the files in the repository, and use replace_in_file to make changes to the files.
+
+However some MCP servers may be running from installed packages rather than a local repository, in which case it may make more sense to create a new MCP server.
+
+# MCP Servers Are Not Always Necessary
+
+The user may not always request the use or creation of MCP servers. Instead, they might provide tasks that can be completed with existing tools. While using the MCP SDK to extend your capabilities can be useful, it's important to understand that this is just one specialized type of task you can accomplish. You should only implement MCP servers when the user explicitly requests it (e.g., "add a tool that...").
+
+Remember: The MCP documentation and example provided above are to help you understand and work with existing MCP servers or create new ones when requested by the user. You already have access to tools and capabilities that can be used to accomplish a wide range of tasks.`
+}

src/core/prompts/responses.ts

@@ -2,6 +2,7 @@ import { Anthropic } from "@anthropic-ai/sdk"
 import * as diff from "diff"
 import * as path from "path"
 import { ClineIgnoreController, LOCK_TEXT_SYMBOL } from "../ignore/ClineIgnoreController"
+import { McpToolCallResponse } from "../../shared/mcp"
 
 export const formatResponse = {
 	duplicateFileReadNotice: () =>
@@ -196,7 +197,7 @@ Otherwise, if you have not completed the task and do not need additional informa
 		`This is likely because the SEARCH block content doesn't match exactly with what's in the file, or if you used multiple SEARCH/REPLACE blocks they may not have been in the order they appear in the file.\n\n` +
 		`The file was reverted to its original state:\n\n` +
 		`<file_content path="${relPath.toPosix()}">\n${originalContent}\n</file_content>\n\n` +
-		`Now that you have the latest state of the file, try the operation again with fewer more precise SEARCH blocks. It may be prudent, especially for large files, to try to limit yourself to 1-3 SEARCH/REPLACE blocks at a time, then wait for the user to respond with the result of the operation before following up with another replace_in_file call to make additional edits.\n(If you run into this error 3 times in a row, you may use the write_to_file tool as a fallback.)`,
+		`Now that you have the latest state of the file, try the operation again with fewer, more precise SEARCH blocks. For large files especially, it may be prudent to try to limit yourself to <5 SEARCH/REPLACE blocks at a time, then wait for the user to respond with the result of the operation before following up with another replace_in_file call to make additional edits.\n(If you run into this error 3 times in a row, you may use the write_to_file tool as a fallback.)`,
 
 	toolAlreadyUsed: (toolName: string) =>
 		`Tool [${toolName}] was not executed because a tool has already been used in this message. Only one tool may be used per message. You must assess the first tool's result before proceeding to use the next tool.`,
@@ -204,10 +205,13 @@ Otherwise, if you have not completed the task and do not need additional informa
 	clineIgnoreInstructions: (content: string) =>
 		`# .clineignore\n\n(The following is provided by a root-level .clineignore file where the user has specified files and directories that should not be accessed. When using list_files, you'll notice a ${LOCK_TEXT_SYMBOL} next to files that are blocked. Attempting to access the file's contents e.g. through read_file will result in an error.)\n\n${content}\n.clineignore`,
 
-	clineRulesDirectoryInstructions: (cwd: string, content: string) =>
+	clineRulesGlobalDirectoryInstructions: (globalClineRulesFilePath: string, content: string) =>
+		`# .clinerules/\n\nThe following is provided by a global .clinerules/ directory, located at ${globalClineRulesFilePath.toPosix()}, where the user has specified instructions for all working directories:\n\n${content}`,
+
+	clineRulesLocalDirectoryInstructions: (cwd: string, content: string) =>
 		`# .clinerules/\n\nThe following is provided by a root-level .clinerules/ directory where the user has specified instructions for this working directory (${cwd.toPosix()})\n\n${content}`,
 
-	clineRulesFileInstructions: (cwd: string, content: string) =>
+	clineRulesLocalFileInstructions: (cwd: string, content: string) =>
 		`# .clinerules\n\nThe following is provided by a root-level .clinerules file where the user has specified instructions for this working directory (${cwd.toPosix()})\n\n${content}`,
 }
 

src/core/prompts/system.ts

@@ -177,9 +177,6 @@ Usage:
 		: ""
 }
 
-${
-	mcpHub.getMode() !== "off"
-		? `
 ## use_mcp_tool
 Description: Request to use a tool provided by a connected MCP server. Each MCP server can provide multiple tools with different capabilities. Tools have defined input schemas that specify required and optional parameters.
 Parameters:
@@ -208,9 +205,6 @@ Usage:
 <server_name>server name here</server_name>
 <uri>resource URI here</uri>
 </access_mcp_resource>
-`
-		: ""
-}
 
 ## ask_followup_question
 Description: Ask the user a question to gather additional information needed to complete the task. This tool should be used when you encounter ambiguities, need clarification, or require more details to proceed effectively. It allows for interactive problem-solving by enabling direct communication with the user. Use this tool judiciously to maintain a balance between gathering necessary information and avoiding excessive back-and-forth.
@@ -300,7 +294,25 @@ Usage:
 </content>
 </write_to_file>
 
-## Example 3: Requesting to make targeted edits to a file
+## Example 3: Creating a new task
+
+<new_task>
+<context>
+Authentication System Implementation:
+- We've implemented the basic user model with email/password
+- Password hashing is working with bcrypt
+- Login endpoint is functional with proper validation
+- JWT token generation is implemented
+
+Next Steps:
+- Implement refresh token functionality
+- Add token validation middleware
+- Create password reset flow
+- Implement role-based access control
+</context>
+</new_task>
+
+## Example 4: Requesting to make targeted edits to a file
 
 <replace_in_file>
 <path>src/components/App.tsx</path>
@@ -334,11 +346,8 @@ return (
 >>>>>>> REPLACE
 </diff>
 </replace_in_file>
-${
-	mcpHub.getMode() !== "off"
-		? `
 
-## Example 4: Requesting to use an MCP tool
+## Example 5: Requesting to use an MCP tool
 
 <use_mcp_tool>
 <server_name>weather-server</server_name>
@@ -351,13 +360,6 @@ ${
 </arguments>
 </use_mcp_tool>
 
-## Example 5: Requesting to access an MCP resource
-
-<access_mcp_resource>
-<server_name>weather-server</server_name>
-<uri>weather://san-francisco/current</uri>
-</access_mcp_resource>
-
 ## Example 6: Another example of using an MCP tool (where the server name is a unique identifier such as a URL)
 
 <use_mcp_tool>
@@ -373,9 +375,7 @@ ${
   "assignees": ["octocat"]
 }
 </arguments>
-</use_mcp_tool>`
-		: ""
-}
+</use_mcp_tool>
 
 # Tool Use Guidelines
 
@@ -398,9 +398,6 @@ It is crucial to proceed step-by-step, waiting for the user's message after each
 
 By waiting for and carefully considering the user's response after each tool use, you can react accordingly and make informed decisions about how to proceed with the task. This iterative process helps ensure the overall success and accuracy of your work.
 
-${
-	mcpHub.getMode() !== "off"
-		? `
 ====
 
 MCP SERVERS
@@ -447,374 +444,6 @@ ${
 				})
 				.join("\n\n")}`
 		: "(No MCP servers currently connected)"
-}`
-		: ""
-}
-
-${
-	mcpHub.getMode() === "full"
-		? `
-## Creating an MCP Server
-
-The user may ask you something along the lines of "add a tool" that does some function, in other words to create an MCP server that provides tools and resources that may connect to external APIs for example. You have the ability to create an MCP server and add it to a configuration file that will then expose the tools and resources for you to use with \`use_mcp_tool\` and \`access_mcp_resource\`.
-
-When creating MCP servers, it's important to understand that they operate in a non-interactive environment. The server cannot initiate OAuth flows, open browser windows, or prompt for user input during runtime. All credentials and authentication tokens must be provided upfront through environment variables in the MCP settings configuration. For example, Spotify's API uses OAuth to get a refresh token for the user, but the MCP server cannot initiate this flow. While you can walk the user through obtaining an application client ID and secret, you may have to create a separate one-time setup script (like get-refresh-token.js) that captures and logs the final piece of the puzzle: the user's refresh token (i.e. you might run the script using execute_command which would open a browser for authentication, and then log the refresh token so that you can see it in the command output for you to use in the MCP settings configuration).
-
-Unless the user specifies otherwise, new MCP servers should be created in: ${await mcpHub.getMcpServersPath()}
-
-### Example MCP Server
-
-For example, if the user wanted to give you the ability to retrieve weather information, you could create an MCP server that uses the OpenWeather API to get weather information, add it to the MCP settings configuration file, and then notice that you now have access to new tools and resources in the system prompt that you might use to show the user your new capabilities.
-
-The following example demonstrates how to build an MCP server that provides weather data functionality. While this example shows how to implement resources, resource templates, and tools, in practice you should prefer using tools since they are more flexible and can handle dynamic parameters. The resource and resource template implementations are included here mainly for demonstration purposes of the different MCP capabilities, but a real weather server would likely just expose tools for fetching weather data. (The following steps are for macOS)
-
-1. Use the \`create-typescript-server\` tool to bootstrap a new project in the default MCP servers directory:
-
-\`\`\`bash
-cd ${await mcpHub.getMcpServersPath()}
-npx @modelcontextprotocol/create-server weather-server
-cd weather-server
-# Install dependencies
-npm install axios
-\`\`\`
-
-This will create a new project with the following structure:
-
-\`\`\`
-weather-server/
-  ├── package.json
-      {
-        ...
-        "type": "module", // added by default, uses ES module syntax (import/export) rather than CommonJS (require/module.exports) (Important to know if you create additional scripts in this server repository like a get-refresh-token.js script)
-        "scripts": {
-          "build": "tsc && node -e \"require('fs').chmodSync('build/index.js', '755')\"",
-          ...
-        }
-        ...
-      }
-  ├── tsconfig.json
-  └── src/
-      └── weather-server/
-          └── index.ts      # Main server implementation
-\`\`\`
-
-2. Replace \`src/index.ts\` with the following:
-
-\`\`\`typescript
-#!/usr/bin/env node
-import { Server } from '@modelcontextprotocol/sdk/server/index.js';
-import { StdioServerTransport } from '@modelcontextprotocol/sdk/server/stdio.js';
-import {
-  CallToolRequestSchema,
-  ErrorCode,
-  ListResourcesRequestSchema,
-  ListResourceTemplatesRequestSchema,
-  ListToolsRequestSchema,
-  McpError,
-  ReadResourceRequestSchema,
-} from '@modelcontextprotocol/sdk/types.js';
-import axios from 'axios';
-
-const API_KEY = process.env.OPENWEATHER_API_KEY; // provided by MCP config
-if (!API_KEY) {
-  throw new Error('OPENWEATHER_API_KEY environment variable is required');
-}
-
-interface OpenWeatherResponse {
-  main: {
-    temp: number;
-    humidity: number;
-  };
-  weather: [{ description: string }];
-  wind: { speed: number };
-  dt_txt?: string;
-}
-
-const isValidForecastArgs = (
-  args: any
-): args is { city: string; days?: number } =>
-  typeof args === 'object' &&
-  args !== null &&
-  typeof args.city === 'string' &&
-  (args.days === undefined || typeof args.days === 'number');
-
-class WeatherServer {
-  private server: Server;
-  private axiosInstance;
-
-  constructor() {
-    this.server = new Server(
-      {
-        name: 'example-weather-server',
-        version: '0.1.0',
-      },
-      {
-        capabilities: {
-          resources: {},
-          tools: {},
-        },
-      }
-    );
-
-    this.axiosInstance = axios.create({
-      baseURL: 'http://api.openweathermap.org/data/2.5',
-      params: {
-        appid: API_KEY,
-        units: 'metric',
-      },
-    });
-
-    this.setupResourceHandlers();
-    this.setupToolHandlers();
-    
-    // Error handling
-    this.server.onerror = (error) => console.error('[MCP Error]', error);
-    process.on('SIGINT', async () => {
-      await this.server.close();
-      process.exit(0);
-    });
-  }
-
-  // MCP Resources represent any kind of UTF-8 encoded data that an MCP server wants to make available to clients, such as database records, API responses, log files, and more. Servers define direct resources with a static URI or dynamic resources with a URI template that follows the format \`[protocol]://[host]/[path]\`.
-  private setupResourceHandlers() {
-    // For static resources, servers can expose a list of resources:
-    this.server.setRequestHandler(ListResourcesRequestSchema, async () => ({
-      resources: [
-        // This is a poor example since you could use the resource template to get the same information but this demonstrates how to define a static resource
-        {
-          uri: \`weather://San Francisco/current\`, // Unique identifier for San Francisco weather resource
-          name: \`Current weather in San Francisco\`, // Human-readable name
-          mimeType: 'application/json', // Optional MIME type
-          // Optional description
-          description:
-            'Real-time weather data for San Francisco including temperature, conditions, humidity, and wind speed',
-        },
-      ],
-    }));
-
-    // For dynamic resources, servers can expose resource templates:
-    this.server.setRequestHandler(
-      ListResourceTemplatesRequestSchema,
-      async () => ({
-        resourceTemplates: [
-          {
-            uriTemplate: 'weather://{city}/current', // URI template (RFC 6570)
-            name: 'Current weather for a given city', // Human-readable name
-            mimeType: 'application/json', // Optional MIME type
-            description: 'Real-time weather data for a specified city', // Optional description
-          },
-        ],
-      })
-    );
-
-    // ReadResourceRequestSchema is used for both static resources and dynamic resource templates
-    this.server.setRequestHandler(
-      ReadResourceRequestSchema,
-      async (request) => {
-        const match = request.params.uri.match(
-          /^weather:\/\/([^/]+)\/current$/
-        );
-        if (!match) {
-          throw new McpError(
-            ErrorCode.InvalidRequest,
-            \`Invalid URI format: \${request.params.uri}\`
-          );
-        }
-        const city = decodeURIComponent(match[1]);
-
-        try {
-          const response = await this.axiosInstance.get(
-            'weather', // current weather
-            {
-              params: { q: city },
-            }
-          );
-
-          return {
-            contents: [
-              {
-                uri: request.params.uri,
-                mimeType: 'application/json',
-                text: JSON.stringify(
-                  {
-                    temperature: response.data.main.temp,
-                    conditions: response.data.weather[0].description,
-                    humidity: response.data.main.humidity,
-                    wind_speed: response.data.wind.speed,
-                    timestamp: new Date().toISOString(),
-                  },
-                  null,
-                  2
-                ),
-              },
-            ],
-          };
-        } catch (error) {
-          if (axios.isAxiosError(error)) {
-            throw new McpError(
-              ErrorCode.InternalError,
-              \`Weather API error: \${
-                error.response?.data.message ?? error.message
-              }\`
-            );
-          }
-          throw error;
-        }
-      }
-    );
-  }
-
-  /* MCP Tools enable servers to expose executable functionality to the system. Through these tools, you can interact with external systems, perform computations, and take actions in the real world.
-   * - Like resources, tools are identified by unique names and can include descriptions to guide their usage. However, unlike resources, tools represent dynamic operations that can modify state or interact with external systems.
-   * - While resources and tools are similar, you should prefer to create tools over resources when possible as they provide more flexibility.
-   */
-  private setupToolHandlers() {
-    this.server.setRequestHandler(ListToolsRequestSchema, async () => ({
-      tools: [
-        {
-          name: 'get_forecast', // Unique identifier
-          description: 'Get weather forecast for a city', // Human-readable description
-          inputSchema: {
-            // JSON Schema for parameters
-            type: 'object',
-            properties: {
-              city: {
-                type: 'string',
-                description: 'City name',
-              },
-              days: {
-                type: 'number',
-                description: 'Number of days (1-5)',
-                minimum: 1,
-                maximum: 5,
-              },
-            },
-            required: ['city'], // Array of required property names
-          },
-        },
-      ],
-    }));
-
-    this.server.setRequestHandler(CallToolRequestSchema, async (request) => {
-      if (request.params.name !== 'get_forecast') {
-        throw new McpError(
-          ErrorCode.MethodNotFound,
-          \`Unknown tool: \${request.params.name}\`
-        );
-      }
-
-      if (!isValidForecastArgs(request.params.arguments)) {
-        throw new McpError(
-          ErrorCode.InvalidParams,
-          'Invalid forecast arguments'
-        );
-      }
-
-      const city = request.params.arguments.city;
-      const days = Math.min(request.params.arguments.days || 3, 5);
-
-      try {
-        const response = await this.axiosInstance.get<{
-          list: OpenWeatherResponse[];
-        }>('forecast', {
-          params: {
-            q: city,
-            cnt: days * 8,
-          },
-        });
-
-        return {
-          content: [
-            {
-              type: 'text',
-              text: JSON.stringify(response.data.list, null, 2),
-            },
-          ],
-        };
-      } catch (error) {
-        if (axios.isAxiosError(error)) {
-          return {
-            content: [
-              {
-                type: 'text',
-                text: \`Weather API error: \${
-                  error.response?.data.message ?? error.message
-                }\`,
-              },
-            ],
-            isError: true,
-          };
-        }
-        throw error;
-      }
-    });
-  }
-
-  async run() {
-    const transport = new StdioServerTransport();
-    await this.server.connect(transport);
-    console.error('Weather MCP server running on stdio');
-  }
-}
-
-const server = new WeatherServer();
-server.run().catch(console.error);
-\`\`\`
-
-(Remember: This is just an example–you may use different dependencies, break the implementation up into multiple files, etc.)
-
-3. Build and compile the executable JavaScript file
-
-\`\`\`bash
-npm run build
-\`\`\`
-
-4. Whenever you need an environment variable such as an API key to configure the MCP server, walk the user through the process of getting the key. For example, they may need to create an account and go to a developer dashboard to generate the key. Provide step-by-step instructions and URLs to make it easy for the user to retrieve the necessary information. Then use the ask_followup_question tool to ask the user for the key, in this case the OpenWeather API key.
-
-5. Install the MCP Server by adding the MCP server configuration to the settings file located at '${await mcpHub.getMcpSettingsFilePath()}'. The settings file may have other MCP servers already configured, so you would read it first and then add your new server to the existing \`mcpServers\` object.
-
-IMPORTANT: Regardless of what else you see in the MCP settings file, you must default any new MCP servers you create to disabled=false and autoApprove=[].
-
-\`\`\`json
-{
-  "mcpServers": {
-    ...,
-    "weather": {
-      "command": "node",
-      "args": ["/path/to/weather-server/build/index.js"],
-      "env": {
-        "OPENWEATHER_API_KEY": "user-provided-api-key"
-      }
-    },
-  }
-}
-\`\`\`
-
-(Note: the user may also ask you to install the MCP server to the Claude desktop app, in which case you would read then modify \`~/Library/Application\ Support/Claude/claude_desktop_config.json\` on macOS for example. It follows the same format of a top level \`mcpServers\` object.)
-
-6. After you have edited the MCP settings configuration file, the system will automatically run all the servers and expose the available tools and resources in the 'Connected MCP Servers' section. (Note: If you encounter a 'not connected' error when testing a newly installed mcp server, a common cause is an incorrect build path in your MCP settings configuration. Since compiled JavaScript files are commonly output to either 'dist/' or 'build/' directories, double-check that the build path in your MCP settings matches where your files are actually being compiled. E.g. If you assumed 'build' as the folder, check tsconfig.json to see if it's using 'dist' instead.)
-
-7. Now that you have access to these new tools and resources, you may suggest ways the user can command you to invoke them - for example, with this new weather tool now available, you can invite the user to ask "what's the weather in San Francisco?"
-
-## Editing MCP Servers
-
-The user may ask to add tools or resources that may make sense to add to an existing MCP server (listed under 'Connected MCP Servers' below: ${
-				mcpHub
-					.getServers()
-					.filter((server) => server.status === "connected")
-					.map((server) => server.name)
-					.join(", ") || "(None running currently)"
-			}, e.g. if it would use the same API. This would be possible if you can locate the MCP server repository on the user's system by looking at the server arguments for a filepath. You might then use list_files and read_file to explore the files in the repository, and use replace_in_file to make changes to the files.
-
-However some MCP servers may be running from installed packages rather than a local repository, in which case it may make more sense to create a new MCP server.
-
-# MCP Servers Are Not Always Necessary
-
-The user may not always request the use or creation of MCP servers. Instead, they might provide tasks that can be completed with existing tools. While using the MCP SDK to extend your capabilities can be useful, it's important to understand that this is just one specialized type of task you can accomplish. You should only implement MCP servers when the user explicitly requests it (e.g., "add a tool that...").
-
-Remember: The MCP documentation and example provided above are to help you understand and work with existing MCP servers or create new ones when requested by the user. You already have access to tools and capabilities that can be used to accomplish a wide range of tasks.
-`
-		: ""
 }
 
 ====
@@ -929,13 +558,7 @@ CAPABILITIES
 		? "\n- You can use the browser_action tool to interact with websites (including html files and locally running development servers) through a Puppeteer-controlled browser when you feel it is necessary in accomplishing the user's task. This tool is particularly useful for web development tasks as it allows you to launch a browser, navigate to pages, interact with elements through clicks and keyboard input, and capture the results through screenshots and console logs. This tool may be useful at key stages of web development tasks-such as after implementing new features, making substantial changes, when troubleshooting issues, or to verify the result of your work. You can analyze the provided screenshots to ensure correct rendering or identify errors, and review console logs for runtime issues.\n	- For example, if asked to add a component to a react website, you might create the necessary files, use execute_command to run the site locally, then use browser_action to launch the browser, navigate to the local server, and verify the component renders & functions correctly before closing the browser."
 		: ""
 }
-${
-	mcpHub.getMode() !== "off"
-		? `
 - You have access to MCP servers that may provide additional tools and resources. Each server may provide different capabilities that you can use to accomplish tasks more effectively.
-`
-		: ""
-}
 
 ====
 
@@ -956,7 +579,7 @@ RULES
 - The user may provide a file's contents directly in their message, in which case you shouldn't use the read_file tool to get the file contents again since you already have it.
 - Your goal is to try to accomplish the user's task, NOT engage in a back and forth conversation.${
 	supportsComputerUse
-		? `\n- The user may ask generic non-development tasks, such as "what\'s the latest news" or "look up the weather in San Diego", in which case you might use the browser_action tool to complete the task if it makes sense to do so, rather than trying to create a website or using curl to answer the question.${mcpHub.getMode() !== "off" ? "However, if an available MCP server tool or resource can be used instead, you should prefer to use it over browser_action." : ""}`
+		? `\n- The user may ask generic non-development tasks, such as "what\'s the latest news" or "look up the weather in San Diego", in which case you might use the browser_action tool to complete the task if it makes sense to do so, rather than trying to create a website or using curl to answer the question. However, if an available MCP server tool or resource can be used instead, you should prefer to use it over browser_action.`
 		: ""
 }
 - NEVER end attempt_completion result with a question or request to engage in further conversation! Formulate the end of your result in a way that is final and does not require further input from the user.
@@ -971,13 +594,7 @@ RULES
 		? " Then if you want to test your work, you might use browser_action to launch the site, wait for the user's response confirming the site was launched along with a screenshot, then perhaps e.g., click a button to test functionality if needed, wait for the user's response confirming the button was clicked along with a screenshot of the new state, before finally closing the browser."
 		: ""
 }
-${
-	mcpHub.getMode() !== "off"
-		? `
 - MCP operations should be used one at a time, similar to other tool usage. Wait for confirmation of success before proceeding with additional operations.
-`
-		: ""
-}
 
 ====
 
@@ -1002,7 +619,8 @@ You accomplish a given task iteratively, breaking it down into clear steps and w
 
 export function addUserInstructions(
 	settingsCustomInstructions?: string,
-	clineRulesFileInstructions?: string,
+	globalClineRulesFileInstructions?: string,
+	localClineRulesFileInstructions?: string,
 	clineIgnoreInstructions?: string,
 	preferredLanguageInstructions?: string,
 ) {
@@ -1013,8 +631,11 @@ export function addUserInstructions(
 	if (settingsCustomInstructions) {
 		customInstructions += settingsCustomInstructions + "\n\n"
 	}
-	if (clineRulesFileInstructions) {
-		customInstructions += clineRulesFileInstructions + "\n\n"
+	if (globalClineRulesFileInstructions) {
+		customInstructions += globalClineRulesFileInstructions + "\n\n"
+	}
+	if (localClineRulesFileInstructions) {
+		customInstructions += localClineRulesFileInstructions + "\n\n"
 	}
 	if (clineIgnoreInstructions) {
 		customInstructions += clineIgnoreInstructions

src/core/storage/disk.ts

@@ -5,6 +5,9 @@ import { Anthropic } from "@anthropic-ai/sdk"
 import { fileExistsAtPath } from "../../utils/fs"
 import { ClineMessage } from "../../shared/ExtensionMessage"
 import { TaskMetadata } from "../context/context-tracking/ContextTrackerTypes"
+import os from "os"
+import { execa } from "execa"
+
 export const GlobalFileNames = {
 	apiConversationHistory: "api_conversation_history.json",
 	contextHistory: "context_history.json",
@@ -15,6 +18,42 @@ export const GlobalFileNames = {
 	taskMetadata: "task_metadata.json",
 }
 
+export async function getDocumentsPath(): Promise<string> {
+	if (process.platform === "win32") {
+		try {
+			const { stdout: docsPath } = await execa("powershell", [
+				"-NoProfile", // Ignore user's PowerShell profile(s)
+				"-Command",
+				"[System.Environment]::GetFolderPath([System.Environment+SpecialFolder]::MyDocuments)",
+			])
+			const trimmedPath = docsPath.trim()
+			if (trimmedPath) {
+				return trimmedPath
+			}
+		} catch (err) {
+			console.error("Failed to retrieve Windows Documents path. Falling back to homedir/Documents.")
+		}
+	} else if (process.platform === "linux") {
+		try {
+			// First check if xdg-user-dir exists
+			await execa("which", ["xdg-user-dir"])
+
+			// If it exists, try to get XDG documents path
+			const { stdout } = await execa("xdg-user-dir", ["DOCUMENTS"])
+			const trimmedPath = stdout.trim()
+			if (trimmedPath) {
+				return trimmedPath
+			}
+		} catch {
+			// Log error but continue to fallback
+			console.error("Failed to retrieve XDG Documents path. Falling back to homedir/Documents.")
+		}
+	}
+
+	// Default fallback for all platforms
+	return path.join(os.homedir(), "Documents")
+}
+
 export async function ensureTaskDirectoryExists(context: vscode.ExtensionContext, taskId: string): Promise<string> {
 	const globalStoragePath = context.globalStorageUri.fsPath
 	const taskDir = path.join(globalStoragePath, "tasks", taskId)
@@ -22,6 +61,34 @@ export async function ensureTaskDirectoryExists(context: vscode.ExtensionContext
 	return taskDir
 }
 
+export async function ensureRulesDirectoryExists(): Promise<string> {
+	const userDocumentsPath = await getDocumentsPath()
+	const clineRulesDir = path.join(userDocumentsPath, "Cline", "Rules")
+	try {
+		await fs.mkdir(clineRulesDir, { recursive: true })
+	} catch (error) {
+		return path.join(os.homedir(), "Documents", "Cline", "Rules") // in case creating a directory in documents fails for whatever reason (e.g. permissions) - this is fine because we will fail gracefully with a path that does not exist
+	}
+	return clineRulesDir
+}
+
+export async function ensureMcpServersDirectoryExists(): Promise<string> {
+	const userDocumentsPath = await getDocumentsPath()
+	const mcpServersDir = path.join(userDocumentsPath, "Cline", "MCP")
+	try {
+		await fs.mkdir(mcpServersDir, { recursive: true })
+	} catch (error) {
+		return "~/Documents/Cline/MCP" // in case creating a directory in documents fails for whatever reason (e.g. permissions) - this is fine since this path is only ever used in the system prompt
+	}
+	return mcpServersDir
+}
+
+export async function ensureSettingsDirectoryExists(context: vscode.ExtensionContext): Promise<string> {
+	const settingsDir = path.join(context.globalStorageUri.fsPath, "settings")
+	await fs.mkdir(settingsDir, { recursive: true })
+	return settingsDir
+}
+
 export async function getSavedApiConversationHistory(
 	context: vscode.ExtensionContext,
 	taskId: string,

src/core/storage/state-keys.ts

@@ -19,6 +19,7 @@ export type SecretKey =
 	| "asksageApiKey"
 	| "xaiApiKey"
 	| "sambanovaApiKey"
+
 export type GlobalStateKey =
 	| "apiProvider"
 	| "apiModelId"
@@ -47,6 +48,7 @@ export type GlobalStateKey =
 	| "openRouterModelInfo"
 	| "openRouterProviderSorting"
 	| "autoApprovalSettings"
+	| "globalClineRulesToggles"
 	| "browserSettings"
 	| "chatSettings"
 	| "vsCodeLmModelSelector"
@@ -71,3 +73,5 @@ export type GlobalStateKey =
 	| "reasoningEffort"
 	| "planActSeparateModelsSetting"
 	| "favoritedModelIds"
+
+export type LocalStateKey = "localClineRulesToggles"

src/core/storage/state.ts

@@ -10,6 +10,7 @@ import { BrowserSettings } from "../../shared/BrowserSettings"
 import { ChatSettings } from "../../shared/ChatSettings"
 import { TelemetrySetting } from "../../shared/TelemetrySetting"
 import { UserInfo } from "../../shared/UserInfo"
+import { ClineRulesToggles } from "../../shared/cline-rules"
 /*
 	Storage
 	https://dev.to/kompotkot/how-to-use-secretstorage-in-your-vscode-extensions-2hco
@@ -121,6 +122,7 @@ export async function getAllExtensionState(context: vscode.ExtensionContext) {
 		sambanovaApiKey,
 		planActSeparateModelsSettingRaw,
 		favoritedModelIds,
+		globalClineRulesToggles,
 	] = await Promise.all([
 		getGlobalState(context, "apiProvider") as Promise<ApiProvider | undefined>,
 		getGlobalState(context, "apiModelId") as Promise<string | undefined>,
@@ -191,6 +193,7 @@ export async function getAllExtensionState(context: vscode.ExtensionContext) {
 		getSecret(context, "sambanovaApiKey") as Promise<string | undefined>,
 		getGlobalState(context, "planActSeparateModelsSetting") as Promise<boolean | undefined>,
 		getGlobalState(context, "favoritedModelIds") as Promise<string[] | undefined>,
+		getGlobalState(context, "globalClineRulesToggles") as Promise<ClineRulesToggles | undefined>,
 	])
 
 	let apiProvider: ApiProvider
@@ -207,6 +210,8 @@ export async function getAllExtensionState(context: vscode.ExtensionContext) {
 		}
 	}
 
+	const localClineRulesToggles = (await getWorkspaceState(context, "localClineRulesToggles")) as ClineRulesToggles
+
 	const o3MiniReasoningEffort = vscode.workspace.getConfiguration("cline.modelSettings.o3Mini").get("reasoningEffort", "medium")
 
 	const mcpMarketplaceEnabled = vscode.workspace.getConfiguration("cline").get<boolean>("mcpMarketplace.enabled", true)
@@ -291,6 +296,8 @@ export async function getAllExtensionState(context: vscode.ExtensionContext) {
 		customInstructions,
 		taskHistory,
 		autoApprovalSettings: autoApprovalSettings || DEFAULT_AUTO_APPROVAL_SETTINGS, // default value can be 0 or empty string
+		globalClineRulesToggles: globalClineRulesToggles || {},
+		localClineRulesToggles: localClineRulesToggles || {},
 		browserSettings: { ...DEFAULT_BROWSER_SETTINGS, ...browserSettings }, // this will ensure that older versions of browserSettings (e.g. before remoteBrowserEnabled was added) are merged with the default values (false for remoteBrowserEnabled)
 		chatSettings: chatSettings || DEFAULT_CHAT_SETTINGS,
 		userInfo,

src/core/task/index.ts

@@ -12,7 +12,6 @@ import { ApiHandler, buildApiHandler } from "../../api"
 import { AnthropicHandler } from "../../api/providers/anthropic"
 import { ClineHandler } from "../../api/providers/cline"
 import { OpenRouterHandler } from "../../api/providers/openrouter"
-import { getContextWindowInfo } from "../context/context-management/context-window-utils"
 import { ApiStream } from "../../api/transform/stream"
 import CheckpointTracker from "../../integrations/checkpoints/CheckpointTracker"
 import { DIFF_VIEW_URI_SCHEME, DiffViewProvider } from "../../integrations/editor/DiffViewProvider"
@@ -55,37 +54,43 @@ import { HistoryItem } from "../../shared/HistoryItem"
 import { DEFAULT_LANGUAGE_SETTINGS, getLanguageKey, LanguageDisplay } from "../../shared/Languages"
 import { ClineAskResponse, ClineCheckpointRestore } from "../../shared/WebviewMessage"
 import { calculateApiCostAnthropic } from "../../utils/cost"
-import { fileExistsAtPath, isDirectory } from "../../utils/fs"
-import { arePathsEqual, getReadablePath } from "../../utils/path"
+import { fileExistsAtPath } from "../../utils/fs"
+import { arePathsEqual, getReadablePath, isLocatedInWorkspace } from "../../utils/path"
 import { fixModelHtmlEscaping, removeInvalidChars } from "../../utils/string"
 import { AssistantMessageContent, parseAssistantMessage, ToolParamName, ToolUseName } from ".././assistant-message"
 import { constructNewFileContent } from ".././assistant-message/diff"
-import { ContextManager } from "../context/context-management/ContextManager"
 import { ClineIgnoreController } from ".././ignore/ClineIgnoreController"
 import { parseMentions } from ".././mentions"
 import { formatResponse } from ".././prompts/responses"
 import { addUserInstructions, SYSTEM_PROMPT } from ".././prompts/system"
+import { getContextWindowInfo } from "../context/context-management/context-window-utils"
 import { FileContextTracker } from "../context/context-tracking/FileContextTracker"
 import { ModelContextTracker } from "../context/context-tracking/ModelContextTracker"
-
 import {
 	checkIsAnthropicContextWindowError,
 	checkIsOpenRouterContextWindowError,
 } from "../context/context-management/context-error-handling"
-
+import { ContextManager } from "../context/context-management/ContextManager"
+import { loadMcpDocumentation } from "../prompts/loadMcpDocumentation"
 import {
+	ensureRulesDirectoryExists,
 	ensureTaskDirectoryExists,
 	getSavedApiConversationHistory,
 	getSavedClineMessages,
 	saveApiConversationHistory,
 	saveClineMessages,
 } from "../storage/disk"
-import { McpHub } from "../../services/mcp/McpHub"
-import WorkspaceTracker from "../../integrations/workspace/WorkspaceTracker"
-import { getClineRules } from "../context/instructions/user-instructions/cline-rules"
+import {
+	getGlobalClineRules,
+	getLocalClineRules,
+	refreshClineRulesToggles,
+} from "../context/instructions/user-instructions/cline-rules"
 import { getGlobalState } from "../storage/state"
+import WorkspaceTracker from "../../integrations/workspace/WorkspaceTracker"
+import { McpHub } from "../../services/mcp/McpHub"
 
-const cwd = vscode.workspace.workspaceFolders?.map((folder) => folder.uri.fsPath).at(0) ?? path.join(os.homedir(), "Desktop") // may or may not exist but fs checking existence would immediately ask for permission which would be bad UX, need to come up with a better solution
+export const cwd =
+	vscode.workspace.workspaceFolders?.map((folder) => folder.uri.fsPath).at(0) ?? path.join(os.homedir(), "Desktop") // may or may not exist but fs checking existence would immediately ask for permission which would be bad UX, need to come up with a better solution
 
 type ToolResponse = string | Array<Anthropic.TextBlockParam | Anthropic.ImageBlockParam>
 type UserContent = Array<Anthropic.ContentBlockParam>
@@ -302,9 +307,13 @@ export class Task {
 		}
 	}
 
-	async restoreCheckpoint(messageTs: number, restoreType: ClineCheckpointRestore) {
-		const messageIndex = this.clineMessages.findIndex((m) => m.ts === messageTs)
+	async restoreCheckpoint(messageTs: number, restoreType: ClineCheckpointRestore, offset?: number) {
+		const messageIndex = this.clineMessages.findIndex((m) => m.ts === messageTs) - (offset || 0)
+		// Find the last message before messageIndex that has a lastCheckpointHash
+		const lastHashIndex = findLastIndex(this.clineMessages.slice(0, messageIndex), (m) => m.lastCheckpointHash !== undefined)
 		const message = this.clineMessages[messageIndex]
+		const lastMessageWithHash = this.clineMessages[lastHashIndex]
+
 		if (!message) {
 			console.error("Message not found", this.clineMessages)
 			return
@@ -337,6 +346,14 @@ export class Task {
 						vscode.window.showErrorMessage("Failed to restore checkpoint: " + errorMessage)
 						didWorkspaceRestoreFail = true
 					}
+				} else if (offset && lastMessageWithHash.lastCheckpointHash && this.checkpointTracker) {
+					try {
+						await this.checkpointTracker.resetHead(lastMessageWithHash.lastCheckpointHash)
+					} catch (error) {
+						const errorMessage = error instanceof Error ? error.message : "Unknown error"
+						vscode.window.showErrorMessage("Failed to restore offsetcheckpoint: " + errorMessage)
+						didWorkspaceRestoreFail = true
+					}
 				}
 				break
 		}
@@ -353,10 +370,10 @@ export class Task {
 					await this.overwriteApiConversationHistory(newConversationHistory)
 
 					// update the context history state
-					// await this.contextManager.truncateContextHistory(
-					// 	message.ts,
-					// 	await ensureTaskDirectoryExists(this.getContext(), this.taskId),
-					// )
+					await this.contextManager.truncateContextHistory(
+						message.ts,
+						await ensureTaskDirectoryExists(this.getContext(), this.taskId),
+					)
 
 					// aggregate deleted api reqs info so we don't lose costs/tokens
 					const deletedMessages = this.clineMessages.slice(messageIndex + 1)
@@ -434,7 +451,7 @@ export class Task {
 			return
 		}
 
-		// TODO: handle if this is called from outside original workspace, in which case we need to show user error message we cant show diff outside of workspace?
+		// TODO: handle if this is called from outside original workspace, in which case we need to show user error message we can't show diff outside of workspace?
 		if (!this.checkpointTracker && !this.checkpointTrackerErrorMessage) {
 			try {
 				this.checkpointTracker = await CheckpointTracker.create(this.taskId, this.context.globalStorageUri.fsPath)
@@ -879,12 +896,12 @@ export class Task {
 		await this.overwriteClineMessages(modifiedClineMessages)
 		this.clineMessages = await getSavedClineMessages(this.getContext(), this.taskId)
 
-		// Now present the cline messages to the user and ask if they want to resume (NOTE: we ran into a bug before where the apiconversationhistory wouldnt be initialized when opening a old task, and it was because we were waiting for resume)
+		// Now present the cline messages to the user and ask if they want to resume (NOTE: we ran into a bug before where the apiconversationhistory wouldn't be initialized when opening a old task, and it was because we were waiting for resume)
 		// This is important in case the user deletes messages without resuming the task first
 		this.apiConversationHistory = await getSavedApiConversationHistory(this.getContext(), this.taskId)
 
 		// load the context history state
-		// await this.contextManager.initializeContextHistory(await ensureTaskDirectoryExists(this.getContext(), this.taskId))
+		await this.contextManager.initializeContextHistory(await ensureTaskDirectoryExists(this.getContext(), this.taskId))
 
 		const lastClineMessage = this.clineMessages
 			.slice()
@@ -905,7 +922,6 @@ export class Task {
 		let responseImages: string[] | undefined
 		if (response === "messageResponse") {
 			await this.say("user_feedback", text, images)
-			await this.saveCheckpoint()
 			responseText = text
 			responseImages = images
 		}
@@ -1044,6 +1060,12 @@ export class Task {
 		})
 
 		if (!isAttemptCompletionMessage) {
+			// ensure we aren't creating a duplicate checkpoint
+			const lastMessage = this.clineMessages.at(-1)
+			if (lastMessage?.say === "checkpoint_created") {
+				return
+			}
+
 			// For non-attempt completion we just say checkpoints
 			await this.say("checkpoint_created")
 			this.checkpointTracker?.commit().then(async (commitHash) => {
@@ -1115,34 +1137,85 @@ export class Task {
 
 		let userFeedback: { text?: string; images?: string[] } | undefined
 		let didContinue = false
-		const sendCommandOutput = async (line: string): Promise<void> => {
+
+		// Chunked terminal output buffering
+		const CHUNK_LINE_COUNT = 20
+		const CHUNK_BYTE_SIZE = 2048 // 2KB
+		const CHUNK_DEBOUNCE_MS = 100
+
+		let outputBuffer: string[] = []
+		let outputBufferSize: number = 0
+		let chunkTimer: NodeJS.Timeout | null = null
+		let chunkEnroute = false
+
+		const flushBuffer = async (force = false) => {
+			if (chunkEnroute || outputBuffer.length === 0) {
+				if (force && !chunkEnroute && outputBuffer.length > 0) {
+					// If force is true and no chunkEnroute, flush anyway
+				} else {
+					return
+				}
+			}
+			const chunk = outputBuffer.join("\n")
+			outputBuffer = []
+			outputBufferSize = 0
+			chunkEnroute = true
 			try {
-				const { response, text, images } = await this.ask("command_output", line)
+				const { response, text, images } = await this.ask("command_output", chunk)
 				if (response === "yesButtonClicked") {
 					// proceed while running
 				} else {
 					userFeedback = { text, images }
 				}
 				didContinue = true
-				process.continue() // continue past the await
+				process.continue()
 			} catch {
-				// This can only happen if this ask promise was ignored, so ignore this error
+				// ask promise was ignored
+			} finally {
+				chunkEnroute = false
+				// If more output accumulated while chunkEnroute, flush again
+				if (outputBuffer.length > 0) {
+					await flushBuffer()
+				}
 			}
 		}
 
+		const scheduleFlush = () => {
+			if (chunkTimer) {
+				clearTimeout(chunkTimer)
+			}
+			chunkTimer = setTimeout(() => flushBuffer(), CHUNK_DEBOUNCE_MS)
+		}
+
 		let result = ""
 		process.on("line", (line) => {
 			result += line + "\n"
+
 			if (!didContinue) {
-				sendCommandOutput(line)
+				outputBuffer.push(line)
+				outputBufferSize += Buffer.byteLength(line, "utf8")
+				// Flush if buffer is large enough
+				if (outputBuffer.length >= CHUNK_LINE_COUNT || outputBufferSize >= CHUNK_BYTE_SIZE) {
+					flushBuffer()
+				} else {
+					scheduleFlush()
+				}
 			} else {
 				this.say("command_output", line)
 			}
 		})
 
 		let completed = false
-		process.once("completed", () => {
+		process.once("completed", async () => {
 			completed = true
+			// Flush any remaining buffered output
+			if (!didContinue && outputBuffer.length > 0) {
+				if (chunkTimer) {
+					clearTimeout(chunkTimer)
+					chunkTimer = null
+				}
+				await flushBuffer(true)
+			}
 		})
 
 		process.once("no_shell_integration", async () => {
@@ -1162,7 +1235,6 @@ export class Task {
 
 		if (userFeedback) {
 			await this.say("user_feedback", userFeedback.text, userFeedback.images)
-			await this.saveCheckpoint()
 			return [
 				true,
 				formatResponse.toolResult(
@@ -1207,7 +1279,7 @@ export class Task {
 					]
 				case "execute_command":
 					return [
-						this.autoApprovalSettings.actions.executeSafeCommands,
+						this.autoApprovalSettings.actions.executeSafeCommands ?? false,
 						this.autoApprovalSettings.actions.executeAllCommands ?? false,
 					]
 				case "browser_action":
@@ -1276,7 +1348,12 @@ export class Task {
 				? `# Preferred Language\n\nSpeak in ${preferredLanguage}.`
 				: ""
 
-		const clineRulesFileInstructions = await getClineRules(cwd)
+		const { globalToggles, localToggles } = await refreshClineRulesToggles(this.getContext(), cwd)
+
+		const globalClineRulesFilePath = await ensureRulesDirectoryExists()
+		const globalClineRulesFileInstructions = await getGlobalClineRules(globalClineRulesFilePath, globalToggles)
+
+		const localClineRulesFileInstructions = await getLocalClineRules(cwd, localToggles)
 
 		const clineIgnoreContent = this.clineIgnoreController.clineIgnoreContent
 		let clineIgnoreInstructions: string | undefined
@@ -1286,27 +1363,27 @@ export class Task {
 
 		if (
 			settingsCustomInstructions ||
-			clineRulesFileInstructions ||
+			globalClineRulesFileInstructions ||
+			localClineRulesFileInstructions ||
 			clineIgnoreInstructions ||
 			preferredLanguageInstructions
 		) {
 			// altering the system prompt mid-task will break the prompt cache, but in the grand scheme this will not change often so it's better to not pollute user messages with it the way we have to with <potentially relevant details>
 			systemPrompt += addUserInstructions(
 				settingsCustomInstructions,
-				clineRulesFileInstructions,
+				globalClineRulesFileInstructions,
+				localClineRulesFileInstructions,
 				clineIgnoreInstructions,
 				preferredLanguageInstructions,
 			)
 		}
-
-		// await
-		const contextManagementMetadata = this.contextManager.getNewContextMessagesAndMetadata(
+		const contextManagementMetadata = await this.contextManager.getNewContextMessagesAndMetadata(
 			this.apiConversationHistory,
 			this.clineMessages,
 			this.api,
 			this.conversationHistoryDeletedRange,
 			previousApiReqIndex,
-			// await ensureTaskDirectoryExists(this.getContext(), this.taskId),
+			await ensureTaskDirectoryExists(this.getContext(), this.taskId),
 		)
 
 		if (contextManagementMetadata.updatedConversationHistoryDeletedRange) {
@@ -1502,6 +1579,8 @@ export class Task {
 							return `[${block.name} for '${block.params.question}']`
 						case "plan_mode_respond":
 							return `[${block.name}]`
+						case "load_mcp_documentation":
+							return `[${block.name}]`
 						case "attempt_completion":
 							return `[${block.name}]`
 						case "new_task":
@@ -1579,7 +1658,6 @@ export class Task {
 						if (text || images?.length) {
 							pushAdditionalToolFeedback(text, images)
 							await this.say("user_feedback", text, images)
-							await this.saveCheckpoint()
 						}
 						this.didRejectTool = true // Prevent further tool uses in this message
 						return false
@@ -1588,7 +1666,6 @@ export class Task {
 						if (text || images?.length) {
 							pushAdditionalToolFeedback(text, images)
 							await this.say("user_feedback", text, images)
-							await this.saveCheckpoint()
 						}
 						return true
 					}
@@ -1662,7 +1739,7 @@ export class Task {
 						if (!accessAllowed) {
 							await this.say("clineignore_error", relPath)
 							pushToolResult(formatResponse.toolError(formatResponse.clineIgnoreError(relPath)))
-							await this.saveCheckpoint()
+
 							break
 						}
 
@@ -1718,7 +1795,6 @@ export class Task {
 									)
 									await this.diffViewProvider.revertChanges()
 									await this.diffViewProvider.reset()
-									await this.saveCheckpoint()
 									break
 								}
 							} else if (content) {
@@ -1749,6 +1825,7 @@ export class Task {
 								tool: fileExists ? "editedExistingFile" : "newFileCreated",
 								path: getReadablePath(cwd, removeClosingTag("path", relPath)),
 								content: diff || content,
+								operationIsLocatedInWorkspace: isLocatedInWorkspace(relPath),
 							}
 
 							if (block.partial) {
@@ -1782,14 +1859,14 @@ export class Task {
 									this.consecutiveMistakeCount++
 									pushToolResult(await this.sayAndCreateMissingParamError("replace_in_file", "diff"))
 									await this.diffViewProvider.reset()
-									await this.saveCheckpoint()
+
 									break
 								}
 								if (block.name === "write_to_file" && !content) {
 									this.consecutiveMistakeCount++
 									pushToolResult(await this.sayAndCreateMissingParamError("write_to_file", "content"))
 									await this.diffViewProvider.reset()
-									await this.saveCheckpoint()
+
 									break
 								}
 
@@ -1812,6 +1889,7 @@ export class Task {
 								const completeMessage = JSON.stringify({
 									...sharedMessageProps,
 									content: diff || content,
+									operationIsLocatedInWorkspace: isLocatedInWorkspace(relPath),
 									// ? formatResponse.createPrettyPatch(
 									// 		relPath,
 									// 		this.diffViewProvider.originalContent,
@@ -1847,7 +1925,6 @@ export class Task {
 										if (text || images?.length) {
 											pushAdditionalToolFeedback(text, images)
 											await this.say("user_feedback", text, images)
-											await this.saveCheckpoint()
 										}
 										this.didRejectTool = true
 										didApprove = false
@@ -1857,14 +1934,12 @@ export class Task {
 										if (text || images?.length) {
 											pushAdditionalToolFeedback(text, images)
 											await this.say("user_feedback", text, images)
-											await this.saveCheckpoint()
 										}
 										telemetryService.captureToolUsage(this.taskId, block.name, false, true)
 									}
 
 									if (!didApprove) {
 										await this.diffViewProvider.revertChanges()
-										await this.saveCheckpoint()
 										break
 									}
 								}
@@ -1925,7 +2000,7 @@ export class Task {
 							await handleError("writing file", error)
 							await this.diffViewProvider.revertChanges()
 							await this.diffViewProvider.reset()
-							await this.saveCheckpoint()
+
 							break
 						}
 					}
@@ -1940,6 +2015,7 @@ export class Task {
 								const partialMessage = JSON.stringify({
 									...sharedMessageProps,
 									content: undefined,
+									operationIsLocatedInWorkspace: isLocatedInWorkspace(relPath),
 								} satisfies ClineSayTool)
 								if (this.shouldAutoApproveToolWithPath(block.name, block.params.path)) {
 									this.removeLastPartialMessageIfExistsWithType("ask", "tool")
@@ -1953,7 +2029,7 @@ export class Task {
 								if (!relPath) {
 									this.consecutiveMistakeCount++
 									pushToolResult(await this.sayAndCreateMissingParamError("read_file", "path"))
-									await this.saveCheckpoint()
+
 									break
 								}
 
@@ -1961,7 +2037,7 @@ export class Task {
 								if (!accessAllowed) {
 									await this.say("clineignore_error", relPath)
 									pushToolResult(formatResponse.toolError(formatResponse.clineIgnoreError(relPath)))
-									await this.saveCheckpoint()
+
 									break
 								}
 
@@ -1970,6 +2046,7 @@ export class Task {
 								const completeMessage = JSON.stringify({
 									...sharedMessageProps,
 									content: absolutePath,
+									operationIsLocatedInWorkspace: isLocatedInWorkspace(relPath),
 								} satisfies ClineSayTool)
 								if (this.shouldAutoApproveToolWithPath(block.name, block.params.path)) {
 									this.removeLastPartialMessageIfExistsWithType("ask", "tool")
@@ -1983,7 +2060,6 @@ export class Task {
 									this.removeLastPartialMessageIfExistsWithType("say", "tool")
 									const didApprove = await askApproval("tool", completeMessage)
 									if (!didApprove) {
-										await this.saveCheckpoint()
 										telemetryService.captureToolUsage(this.taskId, block.name, false, false)
 										break
 									}
@@ -1996,12 +2072,12 @@ export class Task {
 								await this.fileContextTracker.trackFileContext(relPath, "read_tool")
 
 								pushToolResult(content)
-								await this.saveCheckpoint()
+
 								break
 							}
 						} catch (error) {
 							await handleError("reading file", error)
-							await this.saveCheckpoint()
+
 							break
 						}
 					}
@@ -2018,6 +2094,7 @@ export class Task {
 								const partialMessage = JSON.stringify({
 									...sharedMessageProps,
 									content: "",
+									operationIsLocatedInWorkspace: isLocatedInWorkspace(block.params.path),
 								} satisfies ClineSayTool)
 								if (this.shouldAutoApproveToolWithPath(block.name, block.params.path)) {
 									this.removeLastPartialMessageIfExistsWithType("ask", "tool")
@@ -2031,7 +2108,7 @@ export class Task {
 								if (!relDirPath) {
 									this.consecutiveMistakeCount++
 									pushToolResult(await this.sayAndCreateMissingParamError("list_files", "path"))
-									await this.saveCheckpoint()
+
 									break
 								}
 								this.consecutiveMistakeCount = 0
@@ -2049,6 +2126,7 @@ export class Task {
 								const completeMessage = JSON.stringify({
 									...sharedMessageProps,
 									content: result,
+									operationIsLocatedInWorkspace: isLocatedInWorkspace(block.params.path),
 								} satisfies ClineSayTool)
 								if (this.shouldAutoApproveToolWithPath(block.name, block.params.path)) {
 									this.removeLastPartialMessageIfExistsWithType("ask", "tool")
@@ -2062,19 +2140,18 @@ export class Task {
 									this.removeLastPartialMessageIfExistsWithType("say", "tool")
 									const didApprove = await askApproval("tool", completeMessage)
 									if (!didApprove) {
-										await this.saveCheckpoint()
 										telemetryService.captureToolUsage(this.taskId, block.name, false, false)
 										break
 									}
 									telemetryService.captureToolUsage(this.taskId, block.name, false, true)
 								}
 								pushToolResult(result)
-								await this.saveCheckpoint()
+
 								break
 							}
 						} catch (error) {
 							await handleError("listing files", error)
-							await this.saveCheckpoint()
+
 							break
 						}
 					}
@@ -2089,6 +2166,7 @@ export class Task {
 								const partialMessage = JSON.stringify({
 									...sharedMessageProps,
 									content: "",
+									operationIsLocatedInWorkspace: isLocatedInWorkspace(block.params.path),
 								} satisfies ClineSayTool)
 								if (this.shouldAutoApproveToolWithPath(block.name, block.params.path)) {
 									this.removeLastPartialMessageIfExistsWithType("ask", "tool")
@@ -2102,7 +2180,7 @@ export class Task {
 								if (!relDirPath) {
 									this.consecutiveMistakeCount++
 									pushToolResult(await this.sayAndCreateMissingParamError("list_code_definition_names", "path"))
-									await this.saveCheckpoint()
+
 									break
 								}
 
@@ -2117,6 +2195,7 @@ export class Task {
 								const completeMessage = JSON.stringify({
 									...sharedMessageProps,
 									content: result,
+									operationIsLocatedInWorkspace: isLocatedInWorkspace(block.params.path),
 								} satisfies ClineSayTool)
 								if (this.shouldAutoApproveToolWithPath(block.name, block.params.path)) {
 									this.removeLastPartialMessageIfExistsWithType("ask", "tool")
@@ -2130,19 +2209,18 @@ export class Task {
 									this.removeLastPartialMessageIfExistsWithType("say", "tool")
 									const didApprove = await askApproval("tool", completeMessage)
 									if (!didApprove) {
-										await this.saveCheckpoint()
 										telemetryService.captureToolUsage(this.taskId, block.name, false, false)
 										break
 									}
 									telemetryService.captureToolUsage(this.taskId, block.name, false, true)
 								}
 								pushToolResult(result)
-								await this.saveCheckpoint()
+
 								break
 							}
 						} catch (error) {
 							await handleError("parsing source code definitions", error)
-							await this.saveCheckpoint()
+
 							break
 						}
 					}
@@ -2161,6 +2239,7 @@ export class Task {
 								const partialMessage = JSON.stringify({
 									...sharedMessageProps,
 									content: "",
+									operationIsLocatedInWorkspace: isLocatedInWorkspace(block.params.path),
 								} satisfies ClineSayTool)
 								if (this.shouldAutoApproveToolWithPath(block.name, block.params.path)) {
 									this.removeLastPartialMessageIfExistsWithType("ask", "tool")
@@ -2174,13 +2253,13 @@ export class Task {
 								if (!relDirPath) {
 									this.consecutiveMistakeCount++
 									pushToolResult(await this.sayAndCreateMissingParamError("search_files", "path"))
-									await this.saveCheckpoint()
+
 									break
 								}
 								if (!regex) {
 									this.consecutiveMistakeCount++
 									pushToolResult(await this.sayAndCreateMissingParamError("search_files", "regex"))
-									await this.saveCheckpoint()
+
 									break
 								}
 								this.consecutiveMistakeCount = 0
@@ -2197,6 +2276,7 @@ export class Task {
 								const completeMessage = JSON.stringify({
 									...sharedMessageProps,
 									content: results,
+									operationIsLocatedInWorkspace: isLocatedInWorkspace(block.params.path),
 								} satisfies ClineSayTool)
 								if (this.shouldAutoApproveToolWithPath(block.name, block.params.path)) {
 									this.removeLastPartialMessageIfExistsWithType("ask", "tool")
@@ -2210,19 +2290,18 @@ export class Task {
 									this.removeLastPartialMessageIfExistsWithType("say", "tool")
 									const didApprove = await askApproval("tool", completeMessage)
 									if (!didApprove) {
-										await this.saveCheckpoint()
 										telemetryService.captureToolUsage(this.taskId, block.name, false, false)
 										break
 									}
 									telemetryService.captureToolUsage(this.taskId, block.name, false, true)
 								}
 								pushToolResult(results)
-								await this.saveCheckpoint()
+
 								break
 							}
 						} catch (error) {
 							await handleError("searching files", error)
-							await this.saveCheckpoint()
+
 							break
 						}
 					}
@@ -2238,7 +2317,6 @@ export class Task {
 								this.consecutiveMistakeCount++
 								pushToolResult(await this.sayAndCreateMissingParamError("browser_action", "action"))
 								await this.browserSession.closeBrowser()
-								await this.saveCheckpoint()
 							}
 							break
 						}
@@ -2282,7 +2360,7 @@ export class Task {
 										this.consecutiveMistakeCount++
 										pushToolResult(await this.sayAndCreateMissingParamError("browser_action", "url"))
 										await this.browserSession.closeBrowser()
-										await this.saveCheckpoint()
+
 										break
 									}
 									this.consecutiveMistakeCount = 0
@@ -2298,7 +2376,6 @@ export class Task {
 										this.removeLastPartialMessageIfExistsWithType("say", "browser_action_launch")
 										const didApprove = await askApproval("browser_action_launch", url)
 										if (!didApprove) {
-											await this.saveCheckpoint()
 											break
 										}
 									}
@@ -2324,7 +2401,7 @@ export class Task {
 												await this.sayAndCreateMissingParamError("browser_action", "coordinate"),
 											)
 											await this.browserSession.closeBrowser()
-											await this.saveCheckpoint()
+
 											break // can't be within an inner switch
 										}
 									}
@@ -2333,7 +2410,7 @@ export class Task {
 											this.consecutiveMistakeCount++
 											pushToolResult(await this.sayAndCreateMissingParamError("browser_action", "text"))
 											await this.browserSession.closeBrowser()
-											await this.saveCheckpoint()
+
 											break
 										}
 									}
@@ -2382,7 +2459,7 @@ export class Task {
 												browserActionResult.screenshot ? [browserActionResult.screenshot] : [],
 											),
 										)
-										await this.saveCheckpoint()
+
 										break
 									case "close":
 										pushToolResult(
@@ -2390,7 +2467,7 @@ export class Task {
 												`The browser has been closed. You may now proceed to using other tools.`,
 											),
 										)
-										await this.saveCheckpoint()
+
 										break
 								}
 
@@ -2399,7 +2476,7 @@ export class Task {
 						} catch (error) {
 							await this.browserSession.closeBrowser() // if any error occurs, the browser session is terminated
 							await handleError("executing browser action", error)
-							await this.saveCheckpoint()
+
 							break
 						}
 					}
@@ -2411,7 +2488,7 @@ export class Task {
 						try {
 							if (block.partial) {
 								if (this.shouldAutoApproveTool(block.name)) {
-									// since depending on an upcoming parameter, requiresApproval this may become an ask - we cant partially stream a say prematurely. So in this particular case we have to wait for the requiresApproval parameter to be completed before presenting it.
+									// since depending on an upcoming parameter, requiresApproval this may become an ask - we can't partially stream a say prematurely. So in this particular case we have to wait for the requiresApproval parameter to be completed before presenting it.
 									// await this.say(
 									// 	"command",
 									// 	removeClosingTag("command", command),
@@ -2427,7 +2504,7 @@ export class Task {
 								if (!command) {
 									this.consecutiveMistakeCount++
 									pushToolResult(await this.sayAndCreateMissingParamError("execute_command", "command"))
-									await this.saveCheckpoint()
+
 									break
 								}
 								if (!requiresApprovalRaw) {
@@ -2435,7 +2512,7 @@ export class Task {
 									pushToolResult(
 										await this.sayAndCreateMissingParamError("execute_command", "requires_approval"),
 									)
-									await this.saveCheckpoint()
+
 									break
 								}
 								this.consecutiveMistakeCount = 0
@@ -2457,7 +2534,7 @@ export class Task {
 
 								let didAutoApprove = false
 
-								// If the model says this command is safe and auto aproval for safe commands is true, execute the command
+								// If the model says this command is safe and auto approval for safe commands is true, execute the command
 								// If the model says the command is risky, but *BOTH* auto approve settings are true, execute the command
 								const autoApproveResult = this.shouldAutoApproveTool(block.name)
 								const [autoApproveSafe, autoApproveAll] = Array.isArray(autoApproveResult)
@@ -2483,7 +2560,6 @@ export class Task {
 											`${this.shouldAutoApproveTool(block.name) && requiresApprovalPerLLM ? COMMAND_REQ_APP_STRING : ""}`, // ugly hack until we refactor combineCommandSequences
 									)
 									if (!didApprove) {
-										await this.saveCheckpoint()
 										break
 									}
 								}
@@ -2519,7 +2595,7 @@ export class Task {
 							}
 						} catch (error) {
 							await handleError("executing command", error)
-							await this.saveCheckpoint()
+
 							break
 						}
 					}
@@ -2549,13 +2625,13 @@ export class Task {
 								if (!server_name) {
 									this.consecutiveMistakeCount++
 									pushToolResult(await this.sayAndCreateMissingParamError("use_mcp_tool", "server_name"))
-									await this.saveCheckpoint()
+
 									break
 								}
 								if (!tool_name) {
 									this.consecutiveMistakeCount++
 									pushToolResult(await this.sayAndCreateMissingParamError("use_mcp_tool", "tool_name"))
-									await this.saveCheckpoint()
+
 									break
 								}
 								// arguments are optional, but if they are provided they must be valid JSON
@@ -2579,7 +2655,7 @@ export class Task {
 												formatResponse.invalidMcpToolArgumentError(server_name, tool_name),
 											),
 										)
-										await this.saveCheckpoint()
+
 										break
 									}
 								}
@@ -2606,7 +2682,6 @@ export class Task {
 									this.removeLastPartialMessageIfExistsWithType("say", "use_mcp_server")
 									const didApprove = await askApproval("use_mcp_server", completeMessage)
 									if (!didApprove) {
-										await this.saveCheckpoint()
 										break
 									}
 								}
@@ -2615,8 +2690,13 @@ export class Task {
 								await this.say("mcp_server_request_started") // same as browser_action_result
 								const toolResult = await this.mcpHub.callTool(server_name, tool_name, parsedArguments)
 
-								// TODO: add progress indicator and ability to parse images and non-text responses
-								const toolResultPretty =
+								// TODO: add progress indicator
+
+								const toolResultImages =
+									toolResult?.content
+										.filter((item) => item.type === "image")
+										.map((item) => `data:${item.mimeType};base64,${item.data}`) || []
+								let toolResultText =
 									(toolResult?.isError ? "Error:\n" : "") +
 										toolResult?.content
 											.map((item) => {
@@ -2631,8 +2711,21 @@ export class Task {
 											})
 											.filter(Boolean)
 											.join("\n\n") || "(No response)"
-								await this.say("mcp_server_response", toolResultPretty)
-								pushToolResult(formatResponse.toolResult(toolResultPretty))
+								// webview extracts images from the text response to display in the UI
+								const toolResultToDisplay =
+									toolResultText + toolResultImages?.map((image) => `\n\n${image}`).join("")
+								await this.say("mcp_server_response", toolResultToDisplay)
+
+								// MCP's might return images to display to the user, but the model may not support them
+								const supportsImages = this.api.getModel().info.supportsImages ?? false
+								if (toolResultImages.length > 0 && !supportsImages) {
+									toolResultText += `\n\n[${toolResultImages.length} images were provided in the response, and while they are displayed to the user, you do not have the ability to view them.]`
+								}
+
+								// only passes in images if model supports them
+								pushToolResult(
+									formatResponse.toolResult(toolResultText, supportsImages ? toolResultImages : undefined),
+								)
 
 								await this.saveCheckpoint()
 
@@ -2640,7 +2733,7 @@ export class Task {
 							}
 						} catch (error) {
 							await handleError("executing MCP tool", error)
-							await this.saveCheckpoint()
+
 							break
 						}
 					}
@@ -2668,13 +2761,13 @@ export class Task {
 								if (!server_name) {
 									this.consecutiveMistakeCount++
 									pushToolResult(await this.sayAndCreateMissingParamError("access_mcp_resource", "server_name"))
-									await this.saveCheckpoint()
+
 									break
 								}
 								if (!uri) {
 									this.consecutiveMistakeCount++
 									pushToolResult(await this.sayAndCreateMissingParamError("access_mcp_resource", "uri"))
-									await this.saveCheckpoint()
+
 									break
 								}
 								this.consecutiveMistakeCount = 0
@@ -2695,7 +2788,6 @@ export class Task {
 									this.removeLastPartialMessageIfExistsWithType("say", "use_mcp_server")
 									const didApprove = await askApproval("use_mcp_server", completeMessage)
 									if (!didApprove) {
-										await this.saveCheckpoint()
 										break
 									}
 								}
@@ -2715,12 +2807,12 @@ export class Task {
 										.join("\n\n") || "(Empty response)"
 								await this.say("mcp_server_response", resourceResultPretty)
 								pushToolResult(formatResponse.toolResult(resourceResultPretty))
-								await this.saveCheckpoint()
+
 								break
 							}
 						} catch (error) {
 							await handleError("accessing MCP resource", error)
-							await this.saveCheckpoint()
+
 							break
 						}
 					}
@@ -2739,7 +2831,7 @@ export class Task {
 								if (!question) {
 									this.consecutiveMistakeCount++
 									pushToolResult(await this.sayAndCreateMissingParamError("ask_followup_question", "question"))
-									await this.saveCheckpoint()
+
 									break
 								}
 								this.consecutiveMistakeCount = 0
@@ -2776,12 +2868,12 @@ export class Task {
 								}
 
 								pushToolResult(formatResponse.toolResult(`<answer>\n${text}\n</answer>`, images))
-								await this.saveCheckpoint()
+
 								break
 							}
 						} catch (error) {
 							await handleError("asking question", error)
-							await this.saveCheckpoint()
+
 							break
 						}
 					}
@@ -2795,7 +2887,6 @@ export class Task {
 								if (!context) {
 									this.consecutiveMistakeCount++
 									pushToolResult(await this.sayAndCreateMissingParamError("new_task", "context"))
-									await this.saveCheckpoint()
 									break
 								}
 								this.consecutiveMistakeCount = 0
@@ -2824,12 +2915,10 @@ export class Task {
 										formatResponse.toolResult(`The user has created a new task with the provided context.`),
 									)
 								}
-								await this.saveCheckpoint()
 								break
 							}
 						} catch (error) {
 							await handleError("creating new task", error)
-							await this.saveCheckpoint()
 							break
 						}
 					}
@@ -2890,7 +2979,6 @@ export class Task {
 									if (text || images?.length) {
 										telemetryService.captureOptionsIgnored(this.taskId, options.length, "plan")
 										await this.say("user_feedback", text ?? "", images)
-										await this.saveCheckpoint()
 									}
 								}
 
@@ -2918,6 +3006,21 @@ export class Task {
 							break
 						}
 					}
+					case "load_mcp_documentation": {
+						try {
+							if (block.partial) {
+								// shouldn't happen
+								break
+							} else {
+								await this.say("load_mcp_documentation", "", undefined, false)
+								pushToolResult(await loadMcpDocumentation(this.mcpHub))
+								break
+							}
+						} catch (error) {
+							await handleError("loading MCP documentation", error)
+							break
+						}
+					}
 					case "attempt_completion": {
 						/*
 						this.consecutiveMistakeCount = 0
@@ -2973,7 +3076,7 @@ export class Task {
 										)
 									} else {
 										// last message is completion_result
-										// we have command string, which means we have the result as well, so finish it (doesnt have to exist yet)
+										// we have command string, which means we have the result as well, so finish it (doesn't have to exist yet)
 										await this.say("completion_result", removeClosingTag("result", result), undefined, false)
 										await this.saveCheckpoint(true)
 										await addNewChangesFlagToLastCompletionResultMessage()
@@ -3009,7 +3112,7 @@ export class Task {
 								let commandResult: ToolResponse | undefined
 								if (command) {
 									if (lastMessage && lastMessage.ask !== "command") {
-										// havent sent a command message yet so first send completion_result then command
+										// haven't sent a command message yet so first send completion_result then command
 										await this.say("completion_result", result, undefined, false)
 										await this.saveCheckpoint(true)
 										await addNewChangesFlagToLastCompletionResultMessage()
@@ -3022,14 +3125,12 @@ export class Task {
 									// complete command message
 									const didApprove = await askApproval("command", command)
 									if (!didApprove) {
-										await this.saveCheckpoint()
 										break
 									}
 									const [userRejected, execCommandResult] = await this.executeCommandTool(command!)
 									if (userRejected) {
 										this.didRejectTool = true
 										pushToolResult(execCommandResult)
-										await this.saveCheckpoint()
 										break
 									}
 									// user didn't reject, but the command may have output
@@ -3048,7 +3149,6 @@ export class Task {
 									break
 								}
 								await this.say("user_feedback", text ?? "", images)
-								await this.saveCheckpoint()
 
 								const toolResults: (Anthropic.TextBlockParam | Anthropic.ImageBlockParam)[] = []
 								if (commandResult) {
@@ -3077,7 +3177,7 @@ export class Task {
 							}
 						} catch (error) {
 							await handleError("attempting completion", error)
-							await this.saveCheckpoint()
+
 							break
 						}
 					}
@@ -3381,7 +3481,7 @@ export class Task {
 					if (this.didRejectTool) {
 						// userContent has a tool rejection, so interrupt the assistant's response to present the user's feedback
 						assistantMessage += "\n\n[Response interrupted by user feedback]"
-						// this.userMessageContentReady = true // instead of setting this premptively, we allow the present iterator to finish and set userMessageContentReady when its ready
+						// this.userMessageContentReady = true // instead of setting this preemptively, we allow the present iterator to finish and set userMessageContentReady when its ready
 						break
 					}
 
@@ -3436,7 +3536,7 @@ export class Task {
 			partialBlocks.forEach((block) => {
 				block.partial = false
 			})
-			// this.assistantMessageContent.forEach((e) => (e.partial = false)) // cant just do this bc a tool could be in the middle of executing ()
+			// this.assistantMessageContent.forEach((e) => (e.partial = false)) // can't just do this bc a tool could be in the middle of executing ()
 			if (partialBlocks.length > 0) {
 				this.presentAssistantMessage() // if there is content to update then it will complete and update this.userMessageContentReady to true, which we pwaitfor before making the next request. all this is really doing is presenting the last partial message that we just set to complete
 			}

src/core/webview/index.ts

@@ -193,7 +193,7 @@ export class WebviewProvider implements vscode.WebviewViewProvider {
 		/*
 				content security policy of your webview to only allow scripts that have a specific nonce
 				create a content security policy meta tag so that only loading scripts with a nonce is allowed
-				As your extension grows you will likely want to add custom styles, fonts, and/or images to your webview. If you do, you will need to update the content security policy meta tag to explicity allow for these resources. E.g.
+				As your extension grows you will likely want to add custom styles, fonts, and/or images to your webview. If you do, you will need to update the content security policy meta tag to explicitly allow for these resources. E.g.
 								<meta http-equiv="Content-Security-Policy" content="default-src 'none'; style-src ${webview.cspSource}; font-src ${webview.cspSource}; img-src ${webview.cspSource} https:; script-src 'nonce-${nonce}';">
 		- 'unsafe-inline' is required for styles due to vscode-webview-toolkit's dynamic style injection
 		- since we pass base64 images to the webview, we need to specify img-src ${webview.cspSource} data:;