chore(cli): prepare release v0.0.46

apps/cli/CHANGELOG.md

@@ -5,6 +5,34 @@ All notable changes to the `@roo-code/cli` package will be documented in this fi
 The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.0.0/),
 and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
 
+## [0.0.46] - 2026-01-12
+
+### Added
+
+- **Text User Interface (TUI)**: Major new interactive terminal UI with React/Ink for enhanced user experience ([#10480](https://github.com/RooCodeInc/Roo-Code/pull/10480))
+    - Interactive mode and model pickers for easy selection
+    - Improved task management and navigation
+- CLI release script now supports local installation for testing ([#10597](https://github.com/RooCodeInc/Roo-Code/pull/10597))
+
+### Changed
+
+- Default model changed to `anthropic/claude-opus-4.5` ([#10544](https://github.com/RooCodeInc/Roo-Code/pull/10544))
+- File organization improvements for better maintainability ([#10599](https://github.com/RooCodeInc/Roo-Code/pull/10599))
+- Cleanup in ExtensionHost for better code organization ([#10600](https://github.com/RooCodeInc/Roo-Code/pull/10600))
+- Updated README documentation
+- Logging cleanup and improvements
+
+### Fixed
+
+- Model switching issues (model ID mismatch)
+- ACP task cancellation handling
+- Command output streaming
+- Use `DEFAULT_FLAGS.model` as single source of truth for default model ID
+
+### Tests
+
+- Updated tests for model changes
+
 ## [0.0.45] - 2026-01-08
 
 ### Changed

apps/cli/README.md

@@ -157,7 +157,7 @@ Tokens are valid for 90 days. The CLI will prompt you to re-authenticate when yo
 | `-y, --yes`                       | Non-interactive mode: auto-approve all actions                                          | `false`                       |
 | `-k, --api-key <key>`             | API key for the LLM provider                                                            | From env var                  |
 | `-p, --provider <provider>`       | API provider (anthropic, openai, openrouter, etc.)                                      | `openrouter`                  |
-| `-m, --model <model>`             | Model to use                                                                            | `anthropic/claude-sonnet-4.5` |
+| `-m, --model <model>`             | Model to use                                                                            | `anthropic/claude-opus-4.5`   |
 | `-M, --mode <mode>`               | Mode to start in (code, architect, ask, debug, etc.)                                    | `code`                        |
 | `-r, --reasoning-effort <effort>` | Reasoning effort level (unspecified, disabled, none, minimal, low, medium, high, xhigh) | `medium`                      |
 | `--ephemeral`                     | Run without persisting state (uses temporary storage)                                   | `false`                       |
@@ -171,6 +171,56 @@ Tokens are valid for 90 days. The CLI will prompt you to re-authenticate when yo
 | `roo auth logout` | Clear stored authentication token  |
 | `roo auth status` | Show current authentication status |
 
+## ACP (Agent Client Protocol) Integration
+
+The CLI supports the [Agent Client Protocol (ACP)](https://agentclientprotocol.com), allowing ACP-compatible editors like [Zed](https://zed.dev) to use Roo Code as their AI coding assistant.
+
+### Running ACP Server Mode
+
+Start the CLI in ACP server mode:
+
+```bash
+roo acp [options]
+```
+
+**ACP Options:**
+
+| Option                      | Description                                  | Default                       |
+| --------------------------- | -------------------------------------------- | ----------------------------- |
+| `-e, --extension <path>`    | Path to the extension bundle directory       | Auto-detected                 |
+| `-p, --provider <provider>` | API provider (anthropic, openai, openrouter) | `openrouter`                  |
+| `-m, --model <model>`       | Model to use                                 | `anthropic/claude-opus-4.5`   |
+| `-M, --mode <mode>`         | Initial mode (code, architect, ask, debug)   | `code`                        |
+| `-k, --api-key <key>`       | API key for the LLM provider                 | From env var                  |
+
+### Configuring Zed
+
+Add the following to your Zed settings (`settings.json`):
+
+```json
+{
+	"agent_servers": {
+		"Roo Code": {
+			"command": "roo",
+			"args": ["acp"]
+		}
+	}
+}
+```
+
+If you need to specify options:
+
+```json
+{
+	"agent_servers": {
+		"Roo Code": {
+			"command": "roo",
+			"args": ["acp", "-e", "/path/to/extension", "-m", "anthropic/claude-sonnet-4.5"]
+		}
+	}
+}
+```
+
 ## Environment Variables
 
 The CLI will look for API keys in environment variables if not provided via `--api-key`:

apps/cli/docs/AGENT_LOOP.md

@@ -37,11 +37,57 @@ interface ClineMessage {
 | **say** | Informational - agent is telling you something | No            |
 | **ask** | Interactive - agent needs something from you   | Usually yes   |
 
-## The Key Insight
+## The Key Insight: How the CLI Knows When to Prompt
 
-> **The agent loop stops whenever the last message is an `ask` type (with `partial: false`).**
+The CLI doesn't receive any special "waiting" signal from the extension. Instead, it simply **looks at the last message** and asks three questions:
 
-The specific `ask` value tells you exactly what the agent needs.
+### Is the agent waiting for user input?
+
+```
+isWaitingForInput = true when ALL of these are true:
+
+  1. Last message type is "ask" (not "say")
+  2. Last message is NOT partial: true (streaming is complete)
+  3. The ask type is "blocking" (not "command_output")
+```
+
+That's it. No timing. No special signals. Just look at what the last message is.
+
+### Why this works
+
+When the extension needs user input:
+
+- It sends an `ask` message and **blocks** (waits for response)
+- The ask stays as the last message until the CLI responds
+- CLI sees the ask → prompts user → sends response → extension continues
+
+When auto-approval is enabled:
+
+- Extension sends an `ask` message
+- Extension **immediately auto-responds** to its own ask (doesn't wait)
+- New messages quickly follow the ask
+- CLI sees the ask but it's quickly superseded by newer messages
+- State never "settles" at waiting because the extension kept going
+
+### The Simple Logic
+
+```typescript
+function isWaitingForInput(messages) {
+	const lastMessage = messages.at(-1)
+
+	// Still streaming? Not waiting.
+	if (lastMessage?.partial === true) return false
+
+	// Not an ask? Not waiting.
+	if (lastMessage?.type !== "ask") return false
+
+	// Non-blocking ask? Not waiting.
+	if (lastMessage?.ask === "command_output") return false
+
+	// It's a blocking ask that's done streaming → waiting!
+	return true
+}
+```
 
 ## Ask Categories
 
@@ -348,8 +394,8 @@ Example output:
 ## Summary
 
 1. **Agent communicates via `ClineMessage` stream**
-2. **Last message determines state**
-3. **`ask` messages (non-partial) block the agent**
-4. **Ask category determines required action**
+2. **State detection is simple: look at the last message**
+3. **Waiting = last message is a non-partial, blocking `ask`**
+4. **Auto-approval works by the extension auto-responding to its own asks**
 5. **`partial: true` or `api_req_started` without cost = streaming**
 6. **`ExtensionClient` is the single source of truth**

apps/cli/package.json

@@ -1,6 +1,6 @@
 {
 	"name": "@roo-code/cli",
-	"version": "0.0.45",
+	"version": "0.0.46",
 	"description": "Roo Code CLI - Run the Roo Code agent from the command line",
 	"private": true,
 	"type": "module",
@@ -21,6 +21,7 @@
 		"clean": "rimraf dist .turbo"
 	},
 	"dependencies": {
+		"@agentclientprotocol/sdk": "^0.12.0",
 		"@inkjs/ui": "^2.0.0",
 		"@roo-code/core": "workspace:^",
 		"@roo-code/types": "workspace:^",
@@ -33,6 +34,7 @@
 		"p-wait-for": "^5.0.2",
 		"react": "^19.1.0",
 		"superjson": "^2.2.6",
+		"zod": "^4.3.5",
 		"zustand": "^5.0.0"
 	},
 	"devDependencies": {

apps/cli/src/acp/__tests__/agent.test.ts

@@ -0,0 +1,247 @@
+import type * as acp from "@agentclientprotocol/sdk"
+
+import { RooCodeAgent } from "../agent.js"
+import type { AcpSessionOptions } from "../session.js"
+
+vi.mock("@/commands/auth/index.js", () => ({
+	login: vi.fn().mockResolvedValue({ success: true }),
+	logout: vi.fn().mockResolvedValue({ success: true }),
+	status: vi.fn().mockResolvedValue({ authenticated: false }),
+}))
+
+vi.mock("../session.js", () => ({
+	AcpSession: {
+		create: vi.fn().mockResolvedValue({
+			prompt: vi.fn().mockResolvedValue({ stopReason: "end_turn" }),
+			cancel: vi.fn(),
+			setMode: vi.fn(),
+			dispose: vi.fn().mockResolvedValue(undefined),
+			getSessionId: vi.fn().mockReturnValue("test-session-id"),
+		}),
+	},
+}))
+
+describe("RooCodeAgent", () => {
+	let agent: RooCodeAgent
+	let mockConnection: acp.AgentSideConnection
+
+	const defaultOptions: AcpSessionOptions = {
+		extensionPath: "/test/extension",
+		provider: "openrouter",
+		apiKey: "test-key",
+		model: "test-model",
+		mode: "code",
+	}
+
+	beforeEach(() => {
+		mockConnection = {
+			sessionUpdate: vi.fn().mockResolvedValue(undefined),
+			requestPermission: vi.fn().mockResolvedValue({
+				outcome: { outcome: "selected", optionId: "allow" },
+			}),
+			readTextFile: vi.fn().mockResolvedValue({ content: "test content" }),
+			writeTextFile: vi.fn().mockResolvedValue({}),
+			createTerminal: vi.fn(),
+			extMethod: vi.fn(),
+			extNotification: vi.fn(),
+			signal: new AbortController().signal,
+			closed: Promise.resolve(),
+		} as unknown as acp.AgentSideConnection
+
+		agent = new RooCodeAgent(defaultOptions, mockConnection)
+	})
+
+	afterEach(() => {
+		vi.clearAllMocks()
+	})
+
+	describe("initialize", () => {
+		it("should return protocol version and capabilities", async () => {
+			const result = await agent.initialize({
+				protocolVersion: 1,
+			})
+
+			expect(result.protocolVersion).toBeDefined()
+			expect(result.agentCapabilities).toBeDefined()
+			expect(result.agentCapabilities?.loadSession).toBe(false)
+			expect(result.agentCapabilities?.promptCapabilities?.image).toBe(true)
+		})
+
+		it("should return auth methods", async () => {
+			const result = await agent.initialize({
+				protocolVersion: 1,
+			})
+
+			expect(result.authMethods).toBeDefined()
+			expect(result.authMethods).toHaveLength(1)
+
+			const methods = result.authMethods!
+			expect(methods[0]!.id).toBe("roo")
+		})
+
+		it("should store client capabilities", async () => {
+			const clientCapabilities: acp.ClientCapabilities = {
+				fs: {
+					readTextFile: true,
+					writeTextFile: true,
+				},
+			}
+
+			await agent.initialize({
+				protocolVersion: 1,
+				clientCapabilities,
+			})
+
+			// Capabilities should be stored for use in newSession
+			// This is tested indirectly through the session creation
+		})
+	})
+
+	describe("authenticate", () => {
+		it("should throw for invalid auth method", async () => {
+			await expect(
+				agent.authenticate({
+					methodId: "invalid-method",
+				}),
+			).rejects.toThrow()
+		})
+	})
+
+	describe("newSession", () => {
+		it("should create a new session", async () => {
+			const result = await agent.newSession({
+				cwd: "/test/workspace",
+				mcpServers: [],
+			})
+
+			expect(result.sessionId).toBeDefined()
+			expect(typeof result.sessionId).toBe("string")
+		})
+
+		it("should throw auth error when not authenticated and no API key", async () => {
+			// Create agent without API key
+			const agentWithoutKey = new RooCodeAgent({ ...defaultOptions, apiKey: undefined }, mockConnection)
+
+			// Mock environment to not have API key
+			const originalEnv = process.env.OPENROUTER_API_KEY
+			delete process.env.OPENROUTER_API_KEY
+
+			try {
+				await expect(
+					agentWithoutKey.newSession({
+						cwd: "/test/workspace",
+						mcpServers: [],
+					}),
+				).rejects.toThrow()
+			} finally {
+				if (originalEnv) {
+					process.env.OPENROUTER_API_KEY = originalEnv
+				}
+			}
+		})
+	})
+
+	describe("prompt", () => {
+		it("should forward prompt to session", async () => {
+			// Setup
+			const { sessionId } = await agent.newSession({
+				cwd: "/test/workspace",
+				mcpServers: [],
+			})
+
+			// Execute
+			const result = await agent.prompt({
+				sessionId,
+				prompt: [{ type: "text", text: "Hello, world!" }],
+			})
+
+			// Verify
+			expect(result.stopReason).toBe("end_turn")
+		})
+
+		it("should throw for invalid session ID", async () => {
+			await expect(
+				agent.prompt({
+					sessionId: "invalid-session",
+					prompt: [{ type: "text", text: "Hello" }],
+				}),
+			).rejects.toThrow("Session not found")
+		})
+	})
+
+	describe("cancel", () => {
+		it("should cancel session prompt", async () => {
+			// Setup
+			const { sessionId } = await agent.newSession({
+				cwd: "/test/workspace",
+				mcpServers: [],
+			})
+
+			// Execute - should not throw
+			await agent.cancel({ sessionId })
+		})
+
+		it("should handle cancel for non-existent session gracefully", async () => {
+			// Should not throw for invalid session
+			await agent.cancel({ sessionId: "non-existent" })
+		})
+	})
+
+	describe("setSessionMode", () => {
+		it("should set session mode", async () => {
+			// Setup
+			const { sessionId } = await agent.newSession({
+				cwd: "/test/workspace",
+				mcpServers: [],
+			})
+
+			// Execute
+			const result = await agent.setSessionMode({
+				sessionId,
+				modeId: "architect",
+			})
+
+			// Verify
+			expect(result).toEqual({})
+		})
+
+		it("should throw for invalid mode", async () => {
+			// Setup
+			const { sessionId } = await agent.newSession({
+				cwd: "/test/workspace",
+				mcpServers: [],
+			})
+
+			// Execute
+			await expect(
+				agent.setSessionMode({
+					sessionId,
+					modeId: "invalid-mode",
+				}),
+			).rejects.toThrow("Unknown mode")
+		})
+
+		it("should throw for invalid session", async () => {
+			await expect(
+				agent.setSessionMode({
+					sessionId: "invalid-session",
+					modeId: "code",
+				}),
+			).rejects.toThrow("Session not found")
+		})
+	})
+
+	describe("dispose", () => {
+		it("should dispose all sessions", async () => {
+			// Setup
+			await agent.newSession({ cwd: "/test/workspace1", mcpServers: [] })
+			await agent.newSession({ cwd: "/test/workspace2", mcpServers: [] })
+
+			// Execute
+			await agent.dispose()
+
+			// Verify - creating new session should work (sessions map is cleared)
+			// The next newSession would create a fresh session
+		})
+	})
+})

apps/cli/src/acp/__tests__/command-stream.test.ts

@@ -0,0 +1,314 @@
+/**
+ * Tests for CommandStreamManager
+ *
+ * Tests the command output streaming functionality extracted from session.ts.
+ */
+
+import type { ClineMessage } from "@roo-code/types"
+
+import { DeltaTracker } from "../delta-tracker.js"
+import { CommandStreamManager } from "../command-stream.js"
+import { NullLogger } from "../interfaces.js"
+import type { SendUpdateFn } from "../interfaces.js"
+
+describe("CommandStreamManager", () => {
+	let deltaTracker: DeltaTracker
+	let sendUpdate: SendUpdateFn
+	let sentUpdates: Array<Record<string, unknown>>
+	let manager: CommandStreamManager
+
+	beforeEach(() => {
+		deltaTracker = new DeltaTracker()
+		sentUpdates = []
+		sendUpdate = (update) => {
+			sentUpdates.push(update as Record<string, unknown>)
+		}
+		manager = new CommandStreamManager({
+			deltaTracker,
+			sendUpdate,
+			logger: new NullLogger(),
+		})
+	})
+
+	describe("isCommandOutputMessage", () => {
+		it("returns true for command_output say messages", () => {
+			const message: ClineMessage = {
+				type: "say",
+				say: "command_output",
+				ts: Date.now(),
+				text: "output",
+			}
+			expect(manager.isCommandOutputMessage(message)).toBe(true)
+		})
+
+		it("returns false for other say types", () => {
+			const message: ClineMessage = {
+				type: "say",
+				say: "text",
+				ts: Date.now(),
+				text: "hello",
+			}
+			expect(manager.isCommandOutputMessage(message)).toBe(false)
+		})
+
+		it("returns false for ask messages", () => {
+			const message: ClineMessage = {
+				type: "ask",
+				ask: "command",
+				ts: Date.now(),
+				text: "run command",
+			}
+			expect(manager.isCommandOutputMessage(message)).toBe(false)
+		})
+	})
+
+	describe("trackCommand", () => {
+		it("tracks a pending command", () => {
+			manager.trackCommand("call-1", "npm test", 12345)
+			expect(manager.getPendingCommandCount()).toBe(1)
+		})
+
+		it("tracks multiple commands", () => {
+			manager.trackCommand("call-1", "npm test", 12345)
+			manager.trackCommand("call-2", "npm build", 12346)
+			expect(manager.getPendingCommandCount()).toBe(2)
+		})
+
+		it("overwrites command with same ID", () => {
+			manager.trackCommand("call-1", "npm test", 12345)
+			manager.trackCommand("call-1", "npm build", 12346)
+			expect(manager.getPendingCommandCount()).toBe(1)
+		})
+	})
+
+	describe("handleExecutionOutput", () => {
+		it("does nothing without a pending command", () => {
+			manager.handleExecutionOutput("exec-1", "Hello")
+
+			expect(sentUpdates.length).toBe(0)
+		})
+
+		it("sends opening code fence as agent_message_chunk on first output", () => {
+			manager.trackCommand("call-1", "npm test", 12345)
+			manager.handleExecutionOutput("exec-1", "Hello")
+
+			// First message is opening fence, second is the content
+			expect(sentUpdates.length).toBe(2)
+			expect(sentUpdates[0]).toEqual({
+				sessionUpdate: "agent_message_chunk",
+				content: { type: "text", text: "```\n" },
+			})
+			expect(sentUpdates[1]).toEqual({
+				sessionUpdate: "agent_message_chunk",
+				content: { type: "text", text: "Hello" },
+			})
+		})
+
+		it("sends only delta content on subsequent calls (no fence)", () => {
+			manager.trackCommand("call-1", "npm test", 12345)
+			manager.handleExecutionOutput("exec-1", "Hello")
+			sentUpdates.length = 0 // Clear previous updates
+
+			manager.handleExecutionOutput("exec-1", "Hello World")
+
+			// Only the delta " World" is sent, no fence
+			expect(sentUpdates.length).toBe(1)
+			expect(sentUpdates[0]).toEqual({
+				sessionUpdate: "agent_message_chunk",
+				content: { type: "text", text: " World" },
+			})
+		})
+
+		it("tracks code fence by toolCallId not executionId", () => {
+			manager.trackCommand("call-1", "npm test", 12345)
+
+			// First execution stream
+			manager.handleExecutionOutput("exec-1", "First")
+			expect(manager.hasOpenCodeFences()).toBe(true)
+
+			// Second execution stream for same command - no new opening fence since toolCallId already has one
+			manager.handleExecutionOutput("exec-2", "Second")
+
+			// Should still only have one open fence (tracked by toolCallId)
+			expect(manager.hasOpenCodeFences()).toBe(true)
+
+			// Second call should NOT have opening fence since toolCallId already has one
+			// sentUpdates[0] = opening fence, sentUpdates[1] = "First", sentUpdates[2] = "Second"
+			expect(sentUpdates[2]).toEqual({
+				sessionUpdate: "agent_message_chunk",
+				content: { type: "text", text: "Second" },
+			})
+		})
+
+		it("sends streaming output as agent_message_chunk", () => {
+			manager.trackCommand("call-1", "npm test", 12345)
+			manager.handleExecutionOutput("exec-1", "Running...")
+
+			// Opening fence + content
+			const contentUpdate = sentUpdates.find(
+				(u) =>
+					u.sessionUpdate === "agent_message_chunk" && (u.content as { text: string }).text === "Running...",
+			)
+			expect(contentUpdate).toEqual({
+				sessionUpdate: "agent_message_chunk",
+				content: { type: "text", text: "Running..." },
+			})
+		})
+	})
+
+	describe("handleCommandOutput", () => {
+		it("ignores partial messages", () => {
+			const message: ClineMessage = {
+				type: "say",
+				say: "command_output",
+				ts: Date.now(),
+				text: "partial output",
+				partial: true,
+			}
+
+			manager.handleCommandOutput(message)
+			expect(sentUpdates.length).toBe(0)
+		})
+
+		it("sends closing fence and completion when streaming was used", () => {
+			// Track command and open a code fence via execution output
+			manager.trackCommand("call-1", "npm test", 12345)
+			manager.handleExecutionOutput("exec-1", "output")
+			expect(manager.hasOpenCodeFences()).toBe(true)
+
+			sentUpdates.length = 0 // Clear
+
+			const message: ClineMessage = {
+				type: "say",
+				say: "command_output",
+				ts: Date.now(),
+				text: "final output",
+				partial: false,
+			}
+
+			manager.handleCommandOutput(message)
+
+			// First: closing fence as agent_message_chunk
+			expect(sentUpdates[0]).toEqual({
+				sessionUpdate: "agent_message_chunk",
+				content: { type: "text", text: "```\n" },
+			})
+			// Second: tool_call_update with completed status (no content, just rawOutput)
+			expect(sentUpdates[1]).toEqual({
+				sessionUpdate: "tool_call_update",
+				toolCallId: "call-1",
+				status: "completed",
+				rawOutput: { output: "final output" },
+			})
+			expect(manager.hasOpenCodeFences()).toBe(false)
+		})
+
+		it("sends completion update for pending command without streaming", () => {
+			manager.trackCommand("call-1", "npm test", 12345)
+
+			const message: ClineMessage = {
+				type: "say",
+				say: "command_output",
+				ts: Date.now(),
+				text: "Test passed!",
+				partial: false,
+			}
+
+			manager.handleCommandOutput(message)
+
+			// No streaming, so no closing fence - just the completion update
+			const completionUpdate = sentUpdates.find(
+				(u) => u.sessionUpdate === "tool_call_update" && u.status === "completed",
+			)
+			expect(completionUpdate).toEqual({
+				sessionUpdate: "tool_call_update",
+				toolCallId: "call-1",
+				status: "completed",
+				rawOutput: { output: "Test passed!" },
+			})
+		})
+
+		it("removes pending command after completion", () => {
+			manager.trackCommand("call-1", "npm test", 12345)
+			expect(manager.getPendingCommandCount()).toBe(1)
+
+			const message: ClineMessage = {
+				type: "say",
+				say: "command_output",
+				ts: Date.now(),
+				text: "done",
+				partial: false,
+			}
+
+			manager.handleCommandOutput(message)
+			expect(manager.getPendingCommandCount()).toBe(0)
+		})
+
+		it("picks most recent pending command when multiple exist", () => {
+			manager.trackCommand("call-1", "npm test", 12345)
+			manager.trackCommand("call-2", "npm build", 12346) // More recent
+
+			const message: ClineMessage = {
+				type: "say",
+				say: "command_output",
+				ts: Date.now(),
+				text: "done",
+				partial: false,
+			}
+
+			manager.handleCommandOutput(message)
+
+			const completionUpdate = sentUpdates.find((u) => u.sessionUpdate === "tool_call_update")
+			expect((completionUpdate as Record<string, unknown>).toolCallId).toBe("call-2")
+		})
+	})
+
+	describe("reset", () => {
+		it("clears code fence tracking", () => {
+			manager.trackCommand("call-1", "npm test", 12345)
+			manager.handleExecutionOutput("exec-1", "output")
+			expect(manager.hasOpenCodeFences()).toBe(true)
+
+			manager.reset()
+			expect(manager.hasOpenCodeFences()).toBe(false)
+		})
+
+		it("clears pending commands to avoid stale entries", () => {
+			// Pending commands from previous prompts would cause duplicate completion messages
+			manager.trackCommand("call-1", "npm test", 12345)
+			manager.reset()
+			expect(manager.getPendingCommandCount()).toBe(0)
+		})
+	})
+
+	describe("getPendingCommandCount", () => {
+		it("returns 0 when no commands tracked", () => {
+			expect(manager.getPendingCommandCount()).toBe(0)
+		})
+
+		it("returns correct count", () => {
+			manager.trackCommand("call-1", "cmd1", 1)
+			manager.trackCommand("call-2", "cmd2", 2)
+			expect(manager.getPendingCommandCount()).toBe(2)
+		})
+	})
+
+	describe("hasOpenCodeFences", () => {
+		it("returns false initially", () => {
+			expect(manager.hasOpenCodeFences()).toBe(false)
+		})
+
+		it("returns true after execution output with pending command", () => {
+			manager.trackCommand("call-1", "npm test", 12345)
+			manager.handleExecutionOutput("exec-1", "output")
+			expect(manager.hasOpenCodeFences()).toBe(true)
+		})
+
+		it("returns false after reset", () => {
+			manager.trackCommand("call-1", "npm test", 12345)
+			manager.handleExecutionOutput("exec-1", "output")
+			manager.reset()
+			expect(manager.hasOpenCodeFences()).toBe(false)
+		})
+	})
+})

apps/cli/src/acp/__tests__/content-formatter.test.ts

@@ -0,0 +1,298 @@
+/**
+ * Content Formatter Unit Tests
+ *
+ * Tests for the ContentFormatter class.
+ */
+
+import { ContentFormatter, createContentFormatter } from "../content-formatter.js"
+
+describe("ContentFormatter", () => {
+	describe("formatToolResult", () => {
+		const formatter = new ContentFormatter()
+
+		it("should format search results", () => {
+			const content = "Found 5 results.\n\n# src/file.ts\n  1 | match"
+			const result = formatter.formatToolResult("search", content)
+
+			expect(result).toContain("Found 5 results in 1 file")
+			expect(result).toContain("- src/file.ts")
+			expect(result).toMatch(/^```/)
+			expect(result).toMatch(/```$/)
+		})
+
+		it("should format read results", () => {
+			const content = "line1\nline2\nline3"
+			const result = formatter.formatToolResult("read", content)
+
+			expect(result).toContain("line1")
+			expect(result).toContain("line2")
+			expect(result).toContain("line3")
+			expect(result).toMatch(/^```/)
+			expect(result).toMatch(/```$/)
+		})
+
+		it("should return content unchanged for unknown kinds", () => {
+			const content = "some content"
+			const result = formatter.formatToolResult("unknown", content)
+
+			expect(result).toBe(content)
+		})
+	})
+
+	describe("formatSearchResults", () => {
+		const formatter = new ContentFormatter()
+
+		it("should extract file count and result count", () => {
+			const content = "Found 10 results.\n\n# src/a.ts\n  1 | code\n\n# src/b.ts\n  5 | code"
+			const result = formatter.formatSearchResults(content)
+
+			expect(result).toContain("Found 10 results in 2 files")
+		})
+
+		it("should list unique files alphabetically", () => {
+			const content = "Found 3 results.\n\n# src/z.ts\n  1 | a\n\n# src/a.ts\n  2 | b\n\n# src/m.ts\n  3 | c"
+			const result = formatter.formatSearchResults(content)
+
+			const lines = result.split("\n")
+			const fileLines = lines.filter((l) => l.startsWith("- "))
+
+			expect(fileLines[0]).toBe("- src/a.ts")
+			expect(fileLines[1]).toBe("- src/m.ts")
+			expect(fileLines[2]).toBe("- src/z.ts")
+		})
+
+		it("should deduplicate repeated file paths", () => {
+			const content =
+				"Found 5 results.\n\n# src/file.ts\n  1 | a\n\n# src/file.ts\n  5 | b\n\n# src/other.ts\n  10 | c"
+			const result = formatter.formatSearchResults(content)
+
+			expect(result).toContain("in 2 files")
+			expect((result.match(/- src\/file\.ts/g) || []).length).toBe(1)
+		})
+
+		it("should handle no files found", () => {
+			const content = "No results found"
+			const result = formatter.formatSearchResults(content)
+
+			expect(result).toBe("No results found")
+		})
+
+		it("should handle singular result", () => {
+			const content = "Found 1 result.\n\n# src/file.ts\n  1 | match"
+			const result = formatter.formatSearchResults(content)
+
+			expect(result).toContain("Found 1 result in 1 file")
+		})
+
+		it("should handle missing result count", () => {
+			const content = "# src/file.ts\n  1 | match"
+			const result = formatter.formatSearchResults(content)
+
+			expect(result).toContain("Found matches in 1 file")
+		})
+	})
+
+	describe("formatReadResults", () => {
+		it("should return short content unchanged", () => {
+			const formatter = new ContentFormatter({ maxReadLines: 100 })
+			const content = "line1\nline2\nline3"
+			const result = formatter.formatReadResults(content)
+
+			expect(result).toBe(content)
+		})
+
+		it("should truncate long content", () => {
+			const formatter = new ContentFormatter({ maxReadLines: 5 })
+			const lines = Array.from({ length: 10 }, (_, i) => `line${i + 1}`)
+			const content = lines.join("\n")
+			const result = formatter.formatReadResults(content)
+
+			expect(result).toContain("line1")
+			expect(result).toContain("line5")
+			expect(result).not.toContain("line6")
+			expect(result).toContain("... (5 more lines)")
+		})
+
+		it("should handle exactly maxReadLines", () => {
+			const formatter = new ContentFormatter({ maxReadLines: 5 })
+			const lines = Array.from({ length: 5 }, (_, i) => `line${i + 1}`)
+			const content = lines.join("\n")
+			const result = formatter.formatReadResults(content)
+
+			expect(result).toBe(content)
+		})
+
+		it("should use default maxReadLines of 100", () => {
+			const formatter = new ContentFormatter()
+			const lines = Array.from({ length: 105 }, (_, i) => `line${i + 1}`)
+			const content = lines.join("\n")
+			const result = formatter.formatReadResults(content)
+
+			expect(result).toContain("... (5 more lines)")
+		})
+	})
+
+	describe("wrapInCodeBlock", () => {
+		const formatter = new ContentFormatter()
+
+		it("should wrap content in code block", () => {
+			const result = formatter.wrapInCodeBlock("some code")
+
+			expect(result).toBe("```\nsome code\n```")
+		})
+
+		it("should support language specification", () => {
+			const result = formatter.wrapInCodeBlock("const x = 1", "typescript")
+
+			expect(result).toBe("```typescript\nconst x = 1\n```")
+		})
+
+		it("should handle empty content", () => {
+			const result = formatter.wrapInCodeBlock("")
+
+			expect(result).toBe("```\n\n```")
+		})
+
+		it("should handle multiline content", () => {
+			const result = formatter.wrapInCodeBlock("line1\nline2\nline3")
+
+			expect(result).toBe("```\nline1\nline2\nline3\n```")
+		})
+	})
+
+	describe("extractContentFromRawInput", () => {
+		const formatter = new ContentFormatter()
+
+		it("should extract content field", () => {
+			const result = formatter.extractContentFromRawInput({ content: "my content" })
+			expect(result).toBe("my content")
+		})
+
+		it("should extract text field", () => {
+			const result = formatter.extractContentFromRawInput({ text: "my text" })
+			expect(result).toBe("my text")
+		})
+
+		it("should extract result field", () => {
+			const result = formatter.extractContentFromRawInput({ result: "my result" })
+			expect(result).toBe("my result")
+		})
+
+		it("should extract output field", () => {
+			const result = formatter.extractContentFromRawInput({ output: "my output" })
+			expect(result).toBe("my output")
+		})
+
+		it("should extract fileContent field", () => {
+			const result = formatter.extractContentFromRawInput({ fileContent: "my file content" })
+			expect(result).toBe("my file content")
+		})
+
+		it("should extract data field", () => {
+			const result = formatter.extractContentFromRawInput({ data: "my data" })
+			expect(result).toBe("my data")
+		})
+
+		it("should prioritize content over other fields", () => {
+			const result = formatter.extractContentFromRawInput({
+				content: "content value",
+				text: "text value",
+				result: "result value",
+			})
+			expect(result).toBe("content value")
+		})
+
+		it("should return undefined for empty object", () => {
+			const result = formatter.extractContentFromRawInput({})
+			expect(result).toBeUndefined()
+		})
+
+		it("should return undefined for empty string values", () => {
+			const result = formatter.extractContentFromRawInput({ content: "", text: "" })
+			expect(result).toBeUndefined()
+		})
+
+		it("should skip non-string values", () => {
+			const result = formatter.extractContentFromRawInput({
+				content: 123 as unknown as string,
+				text: "valid text",
+			})
+			expect(result).toBe("valid text")
+		})
+	})
+
+	describe("extractFileContent", () => {
+		const formatter = new ContentFormatter()
+
+		it("should use extractContentFromRawInput for non-readFile tools", () => {
+			const result = formatter.extractFileContent({ tool: "list_files", content: "file list" }, "/workspace")
+			expect(result).toBe("file list")
+		})
+
+		it("should return undefined for readFile with no path", () => {
+			const result = formatter.extractFileContent({ tool: "readFile" }, "/workspace")
+			expect(result).toBeUndefined()
+		})
+
+		// Note: actual file reading is tested in integration tests
+	})
+
+	describe("isUserEcho", () => {
+		const formatter = new ContentFormatter()
+
+		it("should return false for null prompt", () => {
+			expect(formatter.isUserEcho("any text", null)).toBe(false)
+		})
+
+		it("should detect exact match", () => {
+			expect(formatter.isUserEcho("hello world", "hello world")).toBe(true)
+		})
+
+		it("should be case insensitive", () => {
+			expect(formatter.isUserEcho("Hello World", "hello world")).toBe(true)
+		})
+
+		it("should handle whitespace differences", () => {
+			expect(formatter.isUserEcho("  hello world  ", "hello world")).toBe(true)
+		})
+
+		it("should detect text contained in prompt (truncated)", () => {
+			expect(formatter.isUserEcho("write a function", "write a function that adds numbers")).toBe(true)
+		})
+
+		it("should detect prompt contained in text (wrapped)", () => {
+			expect(formatter.isUserEcho("User said: write a function here", "write a function")).toBe(true)
+		})
+
+		it("should not match short strings", () => {
+			expect(formatter.isUserEcho("test", "this is a test prompt")).toBe(false)
+		})
+
+		it("should not match completely different text", () => {
+			expect(formatter.isUserEcho("completely different", "original prompt text")).toBe(false)
+		})
+
+		it("should handle empty strings", () => {
+			expect(formatter.isUserEcho("", "prompt")).toBe(false)
+			expect(formatter.isUserEcho("text", "")).toBe(false)
+		})
+	})
+})
+
+describe("createContentFormatter", () => {
+	it("should create a formatter with default config", () => {
+		const formatter = createContentFormatter()
+		expect(formatter).toBeInstanceOf(ContentFormatter)
+	})
+
+	it("should accept custom config", () => {
+		const formatter = createContentFormatter({ maxReadLines: 50 })
+
+		// Test that custom config is used
+		const lines = Array.from({ length: 55 }, (_, i) => `line${i + 1}`)
+		const content = lines.join("\n")
+		const result = formatter.formatReadResults(content)
+
+		expect(result).toContain("... (5 more lines)")
+	})
+})

apps/cli/src/acp/__tests__/delta-tracker.test.ts

@@ -0,0 +1,137 @@
+import { DeltaTracker } from "../delta-tracker.js"
+
+describe("DeltaTracker", () => {
+	let tracker: DeltaTracker
+
+	beforeEach(() => {
+		tracker = new DeltaTracker()
+	})
+
+	describe("getDelta", () => {
+		it("returns full text on first call for a new id", () => {
+			const delta = tracker.getDelta("msg1", "Hello World")
+			expect(delta).toBe("Hello World")
+		})
+
+		it("returns only new content on subsequent calls", () => {
+			tracker.getDelta("msg1", "Hello")
+			const delta = tracker.getDelta("msg1", "Hello World")
+			expect(delta).toBe(" World")
+		})
+
+		it("returns empty string when text unchanged", () => {
+			tracker.getDelta("msg1", "Hello")
+			const delta = tracker.getDelta("msg1", "Hello")
+			expect(delta).toBe("")
+		})
+
+		it("tracks multiple ids independently", () => {
+			tracker.getDelta("msg1", "Hello")
+			tracker.getDelta("msg2", "Goodbye")
+
+			const delta1 = tracker.getDelta("msg1", "Hello World")
+			const delta2 = tracker.getDelta("msg2", "Goodbye World")
+
+			expect(delta1).toBe(" World")
+			expect(delta2).toBe(" World")
+		})
+
+		it("works with numeric ids (timestamps)", () => {
+			const ts1 = 1234567890
+			const ts2 = 1234567891
+
+			tracker.getDelta(ts1, "First message")
+			tracker.getDelta(ts2, "Second message")
+
+			const delta1 = tracker.getDelta(ts1, "First message updated")
+			const delta2 = tracker.getDelta(ts2, "Second message updated")
+
+			expect(delta1).toBe(" updated")
+			expect(delta2).toBe(" updated")
+		})
+
+		it("handles incremental streaming correctly", () => {
+			// Simulate streaming tokens
+			expect(tracker.getDelta("msg", "H")).toBe("H")
+			expect(tracker.getDelta("msg", "He")).toBe("e")
+			expect(tracker.getDelta("msg", "Hel")).toBe("l")
+			expect(tracker.getDelta("msg", "Hell")).toBe("l")
+			expect(tracker.getDelta("msg", "Hello")).toBe("o")
+		})
+	})
+
+	describe("peekDelta", () => {
+		it("returns delta without updating tracking", () => {
+			tracker.getDelta("msg1", "Hello")
+
+			// Peek should show the delta
+			expect(tracker.peekDelta("msg1", "Hello World")).toBe(" World")
+
+			// But tracking should be unchanged, so getDelta still returns full delta
+			expect(tracker.getDelta("msg1", "Hello World")).toBe(" World")
+
+			// Now peek should show empty
+			expect(tracker.peekDelta("msg1", "Hello World")).toBe("")
+		})
+	})
+
+	describe("reset", () => {
+		it("clears all tracking", () => {
+			tracker.getDelta("msg1", "Hello")
+			tracker.getDelta("msg2", "World")
+
+			tracker.reset()
+
+			// After reset, should get full text again
+			expect(tracker.getDelta("msg1", "Hello")).toBe("Hello")
+			expect(tracker.getDelta("msg2", "World")).toBe("World")
+		})
+	})
+
+	describe("resetId", () => {
+		it("clears tracking for specific id only", () => {
+			tracker.getDelta("msg1", "Hello")
+			tracker.getDelta("msg2", "World")
+
+			tracker.resetId("msg1")
+
+			// msg1 should be reset
+			expect(tracker.getDelta("msg1", "Hello")).toBe("Hello")
+			// msg2 should still be tracked
+			expect(tracker.getDelta("msg2", "World")).toBe("")
+		})
+	})
+
+	describe("getPosition", () => {
+		it("returns 0 for untracked ids", () => {
+			expect(tracker.getPosition("unknown")).toBe(0)
+		})
+
+		it("returns current position for tracked ids", () => {
+			tracker.getDelta("msg1", "Hello")
+			expect(tracker.getPosition("msg1")).toBe(5)
+
+			tracker.getDelta("msg1", "Hello World")
+			expect(tracker.getPosition("msg1")).toBe(11)
+		})
+	})
+
+	describe("edge cases", () => {
+		it("handles empty strings", () => {
+			expect(tracker.getDelta("msg1", "")).toBe("")
+			expect(tracker.getDelta("msg1", "Hello")).toBe("Hello")
+		})
+
+		it("handles unicode correctly", () => {
+			tracker.getDelta("msg1", "Hello 👋")
+			const delta = tracker.getDelta("msg1", "Hello 👋 World 🌍")
+			expect(delta).toBe(" World 🌍")
+		})
+
+		it("handles multiline text", () => {
+			tracker.getDelta("msg1", "Line 1\n")
+			const delta = tracker.getDelta("msg1", "Line 1\nLine 2\n")
+			expect(delta).toBe("Line 2\n")
+		})
+	})
+})

apps/cli/src/acp/__tests__/model-service.test.ts

@@ -0,0 +1,233 @@
+/**
+ * Tests for ModelService
+ */
+
+import { ModelService, createModelService } from "../model-service.js"
+import { DEFAULT_MODELS } from "../types.js"
+
+// Mock fetch globally
+const mockFetch = vi.fn()
+global.fetch = mockFetch
+
+describe("ModelService", () => {
+	beforeEach(() => {
+		vi.clearAllMocks()
+		mockFetch.mockReset()
+	})
+
+	describe("constructor", () => {
+		it("should create a ModelService with default options", () => {
+			const service = new ModelService()
+			expect(service).toBeInstanceOf(ModelService)
+		})
+
+		it("should create a ModelService with custom options", () => {
+			const service = new ModelService({
+				apiUrl: "https://custom.api.com",
+				apiKey: "test-key",
+				timeout: 10000,
+			})
+			expect(service).toBeInstanceOf(ModelService)
+		})
+	})
+
+	describe("createModelService factory", () => {
+		it("should create a ModelService instance", () => {
+			const service = createModelService()
+			expect(service).toBeInstanceOf(ModelService)
+		})
+
+		it("should pass options to ModelService", () => {
+			const service = createModelService({
+				apiKey: "test-api-key",
+			})
+			expect(service).toBeInstanceOf(ModelService)
+		})
+	})
+
+	describe("fetchAvailableModels", () => {
+		it("should return cached models on subsequent calls", async () => {
+			const service = new ModelService()
+
+			// First call - should fetch from API
+			mockFetch.mockResolvedValueOnce({
+				ok: true,
+				json: async () => ({
+					object: "list",
+					data: [
+						{ id: "model-1", owned_by: "test" },
+						{ id: "model-2", owned_by: "test" },
+					],
+				}),
+			})
+
+			const firstResult = await service.fetchAvailableModels()
+			expect(mockFetch).toHaveBeenCalledTimes(1)
+
+			// Second call - should use cache
+			const secondResult = await service.fetchAvailableModels()
+			expect(mockFetch).toHaveBeenCalledTimes(1) // No additional fetch
+			expect(secondResult).toEqual(firstResult)
+		})
+
+		it("should return DEFAULT_MODELS when API fails", async () => {
+			const service = new ModelService()
+
+			mockFetch.mockRejectedValueOnce(new Error("Network error"))
+
+			const result = await service.fetchAvailableModels()
+			expect(result).toEqual(DEFAULT_MODELS)
+		})
+
+		it("should return DEFAULT_MODELS when API returns non-OK status", async () => {
+			const service = new ModelService()
+
+			mockFetch.mockResolvedValueOnce({
+				ok: false,
+				status: 500,
+			})
+
+			const result = await service.fetchAvailableModels()
+			expect(result).toEqual(DEFAULT_MODELS)
+		})
+
+		it("should return DEFAULT_MODELS when API returns invalid response", async () => {
+			const service = new ModelService()
+
+			mockFetch.mockResolvedValueOnce({
+				ok: true,
+				json: async () => ({ invalid: "response" }),
+			})
+
+			const result = await service.fetchAvailableModels()
+			expect(result).toEqual(DEFAULT_MODELS)
+		})
+
+		it("should return DEFAULT_MODELS on timeout", async () => {
+			const service = new ModelService({ timeout: 100 })
+
+			// Mock a fetch that never resolves
+			mockFetch.mockImplementationOnce(
+				() =>
+					new Promise((_, reject) => {
+						setTimeout(() => reject(new DOMException("Aborted", "AbortError")), 50)
+					}),
+			)
+
+			const result = await service.fetchAvailableModels()
+			expect(result).toEqual(DEFAULT_MODELS)
+		})
+
+		it("should transform API response to AcpModel format using name and description fields", async () => {
+			const service = new ModelService()
+
+			mockFetch.mockResolvedValueOnce({
+				ok: true,
+				json: async () => ({
+					data: [
+						{
+							id: "anthropic/claude-3-sonnet",
+							name: "Claude 3 Sonnet",
+							description: "A balanced model for most tasks",
+							owned_by: "anthropic",
+						},
+						{
+							id: "openai/gpt-4",
+							name: "GPT-4",
+							description: "OpenAI's flagship model",
+							owned_by: "openai",
+						},
+					],
+				}),
+			})
+
+			const result = await service.fetchAvailableModels()
+
+			// Should include transformed models with name and description from API
+			expect(result).toHaveLength(2)
+			expect(result).toContainEqual({
+				modelId: "anthropic/claude-3-sonnet",
+				name: "Claude 3 Sonnet",
+				description: "A balanced model for most tasks",
+			})
+			expect(result).toContainEqual({
+				modelId: "openai/gpt-4",
+				name: "GPT-4",
+				description: "OpenAI's flagship model",
+			})
+		})
+
+		it("should sort models by model ID", async () => {
+			const service = new ModelService()
+
+			mockFetch.mockResolvedValueOnce({
+				ok: true,
+				json: async () => ({
+					data: [
+						{ id: "openai/gpt-4", name: "GPT-4" },
+						{ id: "anthropic/claude-3-sonnet", name: "Claude 3 Sonnet" },
+						{ id: "google/gemini-pro", name: "Gemini Pro" },
+					],
+				}),
+			})
+
+			const result = await service.fetchAvailableModels()
+
+			// Should be sorted by model ID
+			expect(result[0]!.modelId).toBe("anthropic/claude-3-sonnet")
+			expect(result[1]!.modelId).toBe("google/gemini-pro")
+			expect(result[2]!.modelId).toBe("openai/gpt-4")
+		})
+
+		it("should include Authorization header when apiKey is provided", async () => {
+			const service = new ModelService({ apiKey: "test-api-key" })
+
+			mockFetch.mockResolvedValueOnce({
+				ok: true,
+				json: async () => ({ data: [] }),
+			})
+
+			await service.fetchAvailableModels()
+
+			expect(mockFetch).toHaveBeenCalledWith(
+				expect.any(String),
+				expect.objectContaining({
+					headers: expect.objectContaining({
+						Authorization: "Bearer test-api-key",
+					}),
+				}),
+			)
+		})
+	})
+
+	describe("clearCache", () => {
+		it("should clear the cached models", async () => {
+			const service = new ModelService()
+
+			// First fetch
+			mockFetch.mockResolvedValueOnce({
+				ok: true,
+				json: async () => ({
+					data: [{ id: "model-1" }],
+				}),
+			})
+
+			await service.fetchAvailableModels()
+			expect(mockFetch).toHaveBeenCalledTimes(1)
+
+			// Clear cache
+			service.clearCache()
+
+			// Second fetch - should call API again
+			mockFetch.mockResolvedValueOnce({
+				ok: true,
+				json: async () => ({
+					data: [{ id: "model-2" }],
+				}),
+			})
+
+			await service.fetchAvailableModels()
+			expect(mockFetch).toHaveBeenCalledTimes(2)
+		})
+	})
+})

apps/cli/src/acp/__tests__/plan-translator.test.ts

@@ -0,0 +1,397 @@
+import type { TodoItem } from "@roo-code/types"
+
+import {
+	todoItemToPlanEntry,
+	todoListToPlanUpdate,
+	parseTodoListFromMessage,
+	isTodoListMessage,
+	extractTodoListFromMessage,
+	createPlanUpdateFromMessage,
+	type PriorityConfig,
+} from "../translator/plan-translator.js"
+
+describe("Plan Translator", () => {
+	// ===========================================================================
+	// Test Data
+	// ===========================================================================
+
+	const createTodoItem = (
+		content: string,
+		status: "pending" | "in_progress" | "completed",
+		id?: string,
+	): TodoItem => ({
+		id: id ?? `todo-${Date.now()}`,
+		content,
+		status,
+	})
+
+	// ===========================================================================
+	// todoItemToPlanEntry
+	// ===========================================================================
+
+	describe("todoItemToPlanEntry", () => {
+		it("converts a todo item to a plan entry with default config", () => {
+			const todo = createTodoItem("Implement feature X", "pending")
+			const entry = todoItemToPlanEntry(todo)
+
+			expect(entry).toEqual({
+				content: "Implement feature X",
+				priority: "medium",
+				status: "pending",
+			})
+		})
+
+		it("assigns high priority to in_progress items by default", () => {
+			const todo = createTodoItem("Working on feature", "in_progress")
+			const entry = todoItemToPlanEntry(todo)
+
+			expect(entry.priority).toBe("high")
+			expect(entry.status).toBe("in_progress")
+		})
+
+		it("preserves completed status", () => {
+			const todo = createTodoItem("Done task", "completed")
+			const entry = todoItemToPlanEntry(todo)
+
+			expect(entry.status).toBe("completed")
+		})
+
+		it("respects custom priority config", () => {
+			const todo = createTodoItem("Low priority task", "pending")
+			const config: PriorityConfig = {
+				defaultPriority: "low",
+				prioritizeInProgress: false,
+				prioritizeByOrder: false,
+				highPriorityCount: 3,
+			}
+			const entry = todoItemToPlanEntry(todo, 0, 1, config)
+
+			expect(entry.priority).toBe("low")
+		})
+
+		it("uses order-based priority when enabled", () => {
+			const config: PriorityConfig = {
+				defaultPriority: "medium",
+				prioritizeInProgress: false,
+				prioritizeByOrder: true,
+				highPriorityCount: 2,
+			}
+
+			// First 2 items should be high priority
+			const first = todoItemToPlanEntry(createTodoItem("First", "pending"), 0, 6, config)
+			const second = todoItemToPlanEntry(createTodoItem("Second", "pending"), 1, 6, config)
+			expect(first.priority).toBe("high")
+			expect(second.priority).toBe("high")
+
+			// Items 3-4 (first half) should be medium
+			const third = todoItemToPlanEntry(createTodoItem("Third", "pending"), 2, 6, config)
+			expect(third.priority).toBe("medium")
+
+			// Items past the halfway point should be low
+			const fifth = todoItemToPlanEntry(createTodoItem("Fifth", "pending"), 4, 6, config)
+			expect(fifth.priority).toBe("low")
+		})
+
+		it("prioritizes in_progress over order when both enabled", () => {
+			const config: PriorityConfig = {
+				defaultPriority: "low",
+				prioritizeInProgress: true,
+				prioritizeByOrder: true,
+				highPriorityCount: 1,
+			}
+
+			// Even at the end of the list, in_progress should be high
+			const inProgress = todoItemToPlanEntry(createTodoItem("In progress", "in_progress"), 5, 6, config)
+			expect(inProgress.priority).toBe("high")
+		})
+	})
+
+	// ===========================================================================
+	// todoListToPlanUpdate
+	// ===========================================================================
+
+	describe("todoListToPlanUpdate", () => {
+		it("converts an empty array to a plan with no entries", () => {
+			const update = todoListToPlanUpdate([])
+
+			expect(update).toEqual({
+				sessionUpdate: "plan",
+				entries: [],
+			})
+		})
+
+		it("converts a list of todos to a plan update", () => {
+			const todos: TodoItem[] = [
+				createTodoItem("Task 1", "completed"),
+				createTodoItem("Task 2", "in_progress"),
+				createTodoItem("Task 3", "pending"),
+			]
+			const update = todoListToPlanUpdate(todos)
+
+			expect(update.sessionUpdate).toBe("plan")
+			expect(update.entries).toHaveLength(3)
+			expect(update.entries[0]).toEqual({
+				content: "Task 1",
+				priority: "medium",
+				status: "completed",
+			})
+			expect(update.entries[1]).toEqual({
+				content: "Task 2",
+				priority: "high", // in_progress gets high priority
+				status: "in_progress",
+			})
+			expect(update.entries[2]).toEqual({
+				content: "Task 3",
+				priority: "medium",
+				status: "pending",
+			})
+		})
+
+		it("accepts partial config overrides", () => {
+			const todos = [createTodoItem("Task", "pending")]
+			const update = todoListToPlanUpdate(todos, { defaultPriority: "high" })
+
+			expect(update.entries[0]?.priority).toBe("high")
+		})
+	})
+
+	// ===========================================================================
+	// parseTodoListFromMessage
+	// ===========================================================================
+
+	describe("parseTodoListFromMessage", () => {
+		it("parses valid todo list JSON", () => {
+			const text = JSON.stringify({
+				tool: "updateTodoList",
+				todos: [
+					{ id: "1", content: "Task 1", status: "pending" },
+					{ id: "2", content: "Task 2", status: "completed" },
+				],
+			})
+
+			const result = parseTodoListFromMessage(text)
+
+			expect(result).toEqual([
+				{ id: "1", content: "Task 1", status: "pending" },
+				{ id: "2", content: "Task 2", status: "completed" },
+			])
+		})
+
+		it("returns null for invalid JSON", () => {
+			expect(parseTodoListFromMessage("not json")).toBeNull()
+			expect(parseTodoListFromMessage("{invalid}")).toBeNull()
+		})
+
+		it("returns null for JSON without updateTodoList tool", () => {
+			expect(parseTodoListFromMessage(JSON.stringify({ tool: "other" }))).toBeNull()
+			expect(parseTodoListFromMessage(JSON.stringify({ todos: [] }))).toBeNull()
+		})
+
+		it("returns null for JSON with non-array todos", () => {
+			expect(parseTodoListFromMessage(JSON.stringify({ tool: "updateTodoList", todos: "not array" }))).toBeNull()
+		})
+	})
+
+	// ===========================================================================
+	// isTodoListMessage
+	// ===========================================================================
+
+	describe("isTodoListMessage", () => {
+		it("detects tool ask messages with updateTodoList", () => {
+			const message = {
+				type: "ask",
+				ask: "tool",
+				text: JSON.stringify({ tool: "updateTodoList", todos: [] }),
+			}
+
+			expect(isTodoListMessage(message)).toBe(true)
+		})
+
+		it("detects user_edit_todos say messages", () => {
+			const message = {
+				type: "say",
+				say: "user_edit_todos",
+				text: JSON.stringify({ tool: "updateTodoList", todos: [] }),
+			}
+
+			expect(isTodoListMessage(message)).toBe(true)
+		})
+
+		it("returns false for other ask types", () => {
+			const message = {
+				type: "ask",
+				ask: "command",
+				text: JSON.stringify({ tool: "updateTodoList", todos: [] }),
+			}
+
+			expect(isTodoListMessage(message)).toBe(false)
+		})
+
+		it("returns false for other say types", () => {
+			const message = {
+				type: "say",
+				say: "text",
+				text: JSON.stringify({ tool: "updateTodoList", todos: [] }),
+			}
+
+			expect(isTodoListMessage(message)).toBe(false)
+		})
+
+		it("returns false for messages without text", () => {
+			const message = {
+				type: "ask",
+				ask: "tool",
+			}
+
+			expect(isTodoListMessage(message)).toBe(false)
+		})
+
+		it("returns false for tool messages with other tools", () => {
+			const message = {
+				type: "ask",
+				ask: "tool",
+				text: JSON.stringify({ tool: "read_file", path: "/some/path" }),
+			}
+
+			expect(isTodoListMessage(message)).toBe(false)
+		})
+	})
+
+	// ===========================================================================
+	// extractTodoListFromMessage
+	// ===========================================================================
+
+	describe("extractTodoListFromMessage", () => {
+		it("extracts todos from tool ask message", () => {
+			const todos = [{ id: "1", content: "Task", status: "pending" }]
+			const message = {
+				type: "ask",
+				ask: "tool",
+				text: JSON.stringify({ tool: "updateTodoList", todos }),
+			}
+
+			expect(extractTodoListFromMessage(message)).toEqual(todos)
+		})
+
+		it("extracts todos from user_edit_todos say message", () => {
+			const todos = [{ id: "1", content: "Task", status: "completed" }]
+			const message = {
+				type: "say",
+				say: "user_edit_todos",
+				text: JSON.stringify({ tool: "updateTodoList", todos }),
+			}
+
+			expect(extractTodoListFromMessage(message)).toEqual(todos)
+		})
+
+		it("returns null for non-todo messages", () => {
+			expect(extractTodoListFromMessage({ type: "say", say: "text", text: "Hello" })).toBeNull()
+		})
+
+		it("returns null for messages without text", () => {
+			expect(extractTodoListFromMessage({ type: "ask", ask: "tool" })).toBeNull()
+		})
+	})
+
+	// ===========================================================================
+	// createPlanUpdateFromMessage
+	// ===========================================================================
+
+	describe("createPlanUpdateFromMessage", () => {
+		it("creates plan update from valid todo message", () => {
+			const todos = [
+				{ id: "1", content: "First task", status: "in_progress" as const },
+				{ id: "2", content: "Second task", status: "pending" as const },
+			]
+			const message = {
+				type: "ask",
+				ask: "tool",
+				text: JSON.stringify({ tool: "updateTodoList", todos }),
+			}
+
+			const update = createPlanUpdateFromMessage(message)
+
+			expect(update).not.toBeNull()
+			expect(update?.sessionUpdate).toBe("plan")
+			expect(update?.entries).toHaveLength(2)
+			expect(update?.entries[0]).toEqual({
+				content: "First task",
+				priority: "high", // in_progress
+				status: "in_progress",
+			})
+		})
+
+		it("returns null for non-todo messages", () => {
+			const message = {
+				type: "say",
+				say: "text",
+				text: "Just some text",
+			}
+
+			expect(createPlanUpdateFromMessage(message)).toBeNull()
+		})
+
+		it("returns null for empty todo list", () => {
+			const message = {
+				type: "ask",
+				ask: "tool",
+				text: JSON.stringify({ tool: "updateTodoList", todos: [] }),
+			}
+
+			expect(createPlanUpdateFromMessage(message)).toBeNull()
+		})
+
+		it("accepts custom priority config", () => {
+			const todos = [{ id: "1", content: "Task", status: "pending" as const }]
+			const message = {
+				type: "ask",
+				ask: "tool",
+				text: JSON.stringify({ tool: "updateTodoList", todos }),
+			}
+
+			const update = createPlanUpdateFromMessage(message, { defaultPriority: "low" })
+
+			expect(update?.entries[0]?.priority).toBe("low")
+		})
+	})
+
+	// ===========================================================================
+	// Edge Cases
+	// ===========================================================================
+
+	describe("edge cases", () => {
+		it("handles todos with special characters in content", () => {
+			const todo = createTodoItem('Task with "quotes" and <html>', "pending")
+			const entry = todoItemToPlanEntry(todo)
+
+			expect(entry.content).toBe('Task with "quotes" and <html>')
+		})
+
+		it("handles todos with unicode content", () => {
+			const todo = createTodoItem("Task with emoji 🚀 and unicode ñ", "pending")
+			const entry = todoItemToPlanEntry(todo)
+
+			expect(entry.content).toBe("Task with emoji 🚀 and unicode ñ")
+		})
+
+		it("handles very long content", () => {
+			const longContent = "A".repeat(10000)
+			const todo = createTodoItem(longContent, "pending")
+			const entry = todoItemToPlanEntry(todo)
+
+			expect(entry.content).toBe(longContent)
+		})
+
+		it("handles malformed JSON gracefully", () => {
+			const message = {
+				type: "ask",
+				ask: "tool",
+				text: '{"tool": "updateTodoList", "todos": [{"broken',
+			}
+
+			expect(isTodoListMessage(message)).toBe(false)
+			expect(extractTodoListFromMessage(message)).toBeNull()
+			expect(createPlanUpdateFromMessage(message)).toBeNull()
+		})
+	})
+})

apps/cli/src/acp/__tests__/prompt-state.test.ts

@@ -0,0 +1,373 @@
+/**
+ * Prompt State Machine Unit Tests
+ *
+ * Tests for the PromptStateMachine class.
+ */
+
+import { PromptStateMachine, createPromptStateMachine } from "../prompt-state.js"
+
+describe("PromptStateMachine", () => {
+	describe("initial state", () => {
+		it("should start in idle state", () => {
+			const sm = new PromptStateMachine()
+			expect(sm.getState()).toBe("idle")
+		})
+
+		it("should have null abort signal initially", () => {
+			const sm = new PromptStateMachine()
+			expect(sm.getAbortSignal()).toBeNull()
+		})
+
+		it("should have null prompt text initially", () => {
+			const sm = new PromptStateMachine()
+			expect(sm.getCurrentPromptText()).toBeNull()
+		})
+	})
+
+	describe("canStartPrompt", () => {
+		it("should return true when idle", () => {
+			const sm = new PromptStateMachine()
+			expect(sm.canStartPrompt()).toBe(true)
+		})
+
+		it("should return false when processing", async () => {
+			const sm = new PromptStateMachine()
+			sm.startPrompt("test")
+
+			expect(sm.canStartPrompt()).toBe(false)
+
+			// Clean up
+			sm.complete(true)
+		})
+	})
+
+	describe("isProcessing", () => {
+		it("should return false when idle", () => {
+			const sm = new PromptStateMachine()
+			expect(sm.isProcessing()).toBe(false)
+		})
+
+		it("should return true when processing", () => {
+			const sm = new PromptStateMachine()
+			sm.startPrompt("test")
+
+			expect(sm.isProcessing()).toBe(true)
+
+			// Clean up
+			sm.complete(true)
+		})
+
+		it("should return false after completion", () => {
+			const sm = new PromptStateMachine()
+			sm.startPrompt("test")
+			sm.complete(true)
+
+			expect(sm.isProcessing()).toBe(false)
+		})
+	})
+
+	describe("startPrompt", () => {
+		it("should transition to processing state", () => {
+			const sm = new PromptStateMachine()
+			sm.startPrompt("test prompt")
+
+			expect(sm.getState()).toBe("processing")
+
+			// Clean up
+			sm.complete(true)
+		})
+
+		it("should store the prompt text", () => {
+			const sm = new PromptStateMachine()
+			sm.startPrompt("my test prompt")
+
+			expect(sm.getCurrentPromptText()).toBe("my test prompt")
+
+			// Clean up
+			sm.complete(true)
+		})
+
+		it("should create abort signal", () => {
+			const sm = new PromptStateMachine()
+			sm.startPrompt("test")
+
+			expect(sm.getAbortSignal()).not.toBeNull()
+			expect(sm.getAbortSignal()?.aborted).toBe(false)
+
+			// Clean up
+			sm.complete(true)
+		})
+
+		it("should return a promise", () => {
+			const sm = new PromptStateMachine()
+			const promise = sm.startPrompt("test")
+
+			expect(promise).toBeInstanceOf(Promise)
+
+			// Clean up
+			sm.complete(true)
+		})
+
+		it("should resolve with end_turn on successful completion", async () => {
+			const sm = new PromptStateMachine()
+			const promise = sm.startPrompt("test")
+
+			sm.complete(true)
+			const result = await promise
+
+			expect(result.stopReason).toBe("end_turn")
+		})
+
+		it("should resolve with refusal on failed completion", async () => {
+			const sm = new PromptStateMachine()
+			const promise = sm.startPrompt("test")
+
+			sm.complete(false)
+			const result = await promise
+
+			expect(result.stopReason).toBe("refusal")
+		})
+
+		it("should resolve with cancelled on cancel", async () => {
+			const sm = new PromptStateMachine()
+			const promise = sm.startPrompt("test")
+
+			sm.cancel()
+			const result = await promise
+
+			expect(result.stopReason).toBe("cancelled")
+		})
+
+		it("should cancel existing prompt if called while processing", async () => {
+			const sm = new PromptStateMachine()
+			const promise1 = sm.startPrompt("first prompt")
+
+			// Start a second prompt (should cancel first)
+			sm.startPrompt("second prompt")
+
+			// First promise should resolve with cancelled
+			const result1 = await promise1
+			expect(result1.stopReason).toBe("cancelled")
+
+			// Clean up
+			sm.complete(true)
+		})
+	})
+
+	describe("complete", () => {
+		it("should transition to idle state", () => {
+			const sm = new PromptStateMachine()
+			sm.startPrompt("test")
+			sm.complete(true)
+
+			expect(sm.getState()).toBe("idle")
+		})
+
+		it("should return end_turn for success", () => {
+			const sm = new PromptStateMachine()
+			sm.startPrompt("test")
+
+			const stopReason = sm.complete(true)
+			expect(stopReason).toBe("end_turn")
+		})
+
+		it("should return refusal for failure", () => {
+			const sm = new PromptStateMachine()
+			sm.startPrompt("test")
+
+			const stopReason = sm.complete(false)
+			expect(stopReason).toBe("refusal")
+		})
+
+		it("should clear prompt text", () => {
+			const sm = new PromptStateMachine()
+			sm.startPrompt("test")
+			sm.complete(true)
+
+			expect(sm.getCurrentPromptText()).toBeNull()
+		})
+
+		it("should clear abort controller", () => {
+			const sm = new PromptStateMachine()
+			sm.startPrompt("test")
+			sm.complete(true)
+
+			expect(sm.getAbortSignal()).toBeNull()
+		})
+
+		it("should be idempotent (multiple calls ignored)", () => {
+			const sm = new PromptStateMachine()
+			sm.startPrompt("test")
+
+			const result1 = sm.complete(true)
+			const result2 = sm.complete(false) // Should be ignored
+
+			expect(result1).toBe("end_turn")
+			expect(result2).toBe("refusal") // Returns mapped value but doesn't change state
+			expect(sm.getState()).toBe("idle")
+		})
+	})
+
+	describe("cancel", () => {
+		it("should abort the signal", () => {
+			const sm = new PromptStateMachine()
+			sm.startPrompt("test")
+			const signal = sm.getAbortSignal()
+
+			sm.cancel()
+
+			expect(signal?.aborted).toBe(true)
+		})
+
+		it("should transition to idle state", async () => {
+			const sm = new PromptStateMachine()
+			const promise = sm.startPrompt("test")
+
+			sm.cancel()
+			await promise
+
+			expect(sm.getState()).toBe("idle")
+		})
+
+		it("should be safe to call when idle", () => {
+			const sm = new PromptStateMachine()
+
+			// Should not throw
+			expect(() => sm.cancel()).not.toThrow()
+			expect(sm.getState()).toBe("idle")
+		})
+
+		it("should be idempotent (multiple calls safe)", () => {
+			const sm = new PromptStateMachine()
+			sm.startPrompt("test")
+
+			sm.cancel()
+			sm.cancel() // Should not throw
+
+			expect(sm.getState()).toBe("idle")
+		})
+	})
+
+	describe("reset", () => {
+		it("should transition to idle state", () => {
+			const sm = new PromptStateMachine()
+			sm.startPrompt("test")
+			sm.reset()
+
+			expect(sm.getState()).toBe("idle")
+		})
+
+		it("should clear prompt text", () => {
+			const sm = new PromptStateMachine()
+			sm.startPrompt("test")
+			sm.reset()
+
+			expect(sm.getCurrentPromptText()).toBeNull()
+		})
+
+		it("should clear abort controller", () => {
+			const sm = new PromptStateMachine()
+			sm.startPrompt("test")
+			sm.reset()
+
+			expect(sm.getAbortSignal()).toBeNull()
+		})
+
+		it("should abort any pending operation", () => {
+			const sm = new PromptStateMachine()
+			sm.startPrompt("test")
+			const signal = sm.getAbortSignal()
+
+			sm.reset()
+
+			expect(signal?.aborted).toBe(true)
+		})
+
+		it("should be safe to call when idle", () => {
+			const sm = new PromptStateMachine()
+
+			expect(() => sm.reset()).not.toThrow()
+			expect(sm.getState()).toBe("idle")
+		})
+	})
+
+	describe("abort signal integration", () => {
+		it("should trigger abort handler on cancel", async () => {
+			const sm = new PromptStateMachine()
+			const promise = sm.startPrompt("test")
+
+			let abortHandlerCalled = false
+			sm.getAbortSignal()?.addEventListener("abort", () => {
+				abortHandlerCalled = true
+			})
+
+			sm.cancel()
+			await promise
+
+			expect(abortHandlerCalled).toBe(true)
+		})
+
+		it("should resolve promise via abort handler", async () => {
+			const sm = new PromptStateMachine()
+			const promise = sm.startPrompt("test")
+
+			sm.cancel()
+			const result = await promise
+
+			expect(result.stopReason).toBe("cancelled")
+		})
+	})
+
+	describe("lifecycle scenarios", () => {
+		it("should handle multiple prompt cycles", async () => {
+			const sm = new PromptStateMachine()
+
+			// First cycle
+			const promise1 = sm.startPrompt("prompt 1")
+			expect(sm.isProcessing()).toBe(true)
+			sm.complete(true)
+			const result1 = await promise1
+			expect(result1.stopReason).toBe("end_turn")
+			expect(sm.isProcessing()).toBe(false)
+
+			// Second cycle
+			const promise2 = sm.startPrompt("prompt 2")
+			expect(sm.isProcessing()).toBe(true)
+			expect(sm.getCurrentPromptText()).toBe("prompt 2")
+			sm.complete(false)
+			const result2 = await promise2
+			expect(result2.stopReason).toBe("refusal")
+
+			// Third cycle with cancellation
+			const promise3 = sm.startPrompt("prompt 3")
+			sm.cancel()
+			const result3 = await promise3
+			expect(result3.stopReason).toBe("cancelled")
+		})
+
+		it("should handle rapid start/cancel cycles", async () => {
+			const sm = new PromptStateMachine()
+
+			const promises: Promise<{ stopReason: string }>[] = []
+
+			for (let i = 0; i < 5; i++) {
+				const promise = sm.startPrompt(`prompt ${i}`)
+				promises.push(promise)
+				sm.cancel()
+			}
+
+			// All should resolve with cancelled
+			const results = await Promise.all(promises)
+			expect(results.every((r) => r.stopReason === "cancelled")).toBe(true)
+		})
+	})
+})
+
+describe("createPromptStateMachine", () => {
+	it("should create a new state machine", () => {
+		const sm = createPromptStateMachine()
+
+		expect(sm).toBeInstanceOf(PromptStateMachine)
+		expect(sm.getState()).toBe("idle")
+	})
+})

apps/cli/src/acp/__tests__/session-plan-integration.test.ts

@@ -0,0 +1,397 @@
+/**
+ * Integration tests for ACP Plan updates via session-event-handler.
+ *
+ * Tests the end-to-end flow of:
+ * 1. Extension sending todo list update messages
+ * 2. Session-event-handler detecting and translating them
+ * 3. ACP plan updates being sent to the connection
+ */
+
+import type { ClineMessage } from "@roo-code/types"
+
+import {
+	SessionEventHandler,
+	createSessionEventHandler,
+	type SessionEventHandlerDeps,
+} from "../session-event-handler.js"
+import type { IAcpLogger, IDeltaTracker, IPromptStateMachine } from "../interfaces.js"
+import { ToolHandlerRegistry } from "../tool-handler.js"
+
+// =============================================================================
+// Mock Setup
+// =============================================================================
+
+const createMockLogger = (): IAcpLogger => ({
+	debug: vi.fn(),
+	info: vi.fn(),
+	warn: vi.fn(),
+	error: vi.fn(),
+	request: vi.fn(),
+	response: vi.fn(),
+	notification: vi.fn(),
+})
+
+const createMockDeltaTracker = (): IDeltaTracker => ({
+	getDelta: vi.fn().mockReturnValue(null),
+	peekDelta: vi.fn().mockReturnValue(null),
+	reset: vi.fn(),
+	resetId: vi.fn(),
+})
+
+const createMockPromptState = (): IPromptStateMachine => ({
+	getState: vi.fn().mockReturnValue("processing"),
+	getAbortSignal: vi.fn().mockReturnValue(null),
+	getPromptText: vi.fn().mockReturnValue(""),
+	canStartPrompt: vi.fn().mockReturnValue(false),
+	isProcessing: vi.fn().mockReturnValue(true), // Return true so messages are processed
+	startPrompt: vi.fn().mockReturnValue(Promise.resolve({ stopReason: "end_turn" })),
+	complete: vi.fn().mockReturnValue("end_turn"),
+	transitionToComplete: vi.fn(),
+	cancel: vi.fn(),
+	reset: vi.fn(),
+})
+
+const createMockCommandStreamManager = () => ({
+	handleExecutionOutput: vi.fn(),
+	handleCommandOutput: vi.fn(),
+	isCommandOutputMessage: vi.fn().mockReturnValue(false),
+	trackCommand: vi.fn(),
+	reset: vi.fn(),
+})
+
+const createMockToolContentStreamManager = () => ({
+	handleToolContentStreaming: vi.fn(),
+	isToolAskMessage: vi.fn().mockReturnValue(false),
+	reset: vi.fn(),
+})
+
+const createMockExtensionClient = () => {
+	const handlers: Record<string, ((data: unknown) => void)[]> = {}
+	return {
+		on: vi.fn((event: string, handler: (data: unknown) => void) => {
+			handlers[event] = handlers[event] || []
+			handlers[event]!.push(handler)
+			return { on: vi.fn(), off: vi.fn() }
+		}),
+		off: vi.fn(),
+		emit: (event: string, data: unknown) => {
+			handlers[event]?.forEach((h) => h(data))
+		},
+		respond: vi.fn(),
+		approve: vi.fn(),
+		reject: vi.fn(),
+	}
+}
+
+const createMockExtensionHost = () => ({
+	on: vi.fn(),
+	off: vi.fn(),
+	client: createMockExtensionClient(),
+	activate: vi.fn().mockResolvedValue(undefined),
+	dispose: vi.fn().mockResolvedValue(undefined),
+	sendToExtension: vi.fn(),
+})
+
+// =============================================================================
+// Tests
+// =============================================================================
+
+describe("Session Plan Integration", () => {
+	let eventHandler: SessionEventHandler
+	let mockSendUpdate: ReturnType<typeof vi.fn>
+	let mockClient: ReturnType<typeof createMockExtensionClient>
+	let deps: SessionEventHandlerDeps
+
+	beforeEach(() => {
+		mockSendUpdate = vi.fn()
+		mockClient = createMockExtensionClient()
+
+		deps = {
+			logger: createMockLogger(),
+			client: mockClient,
+			extensionHost: createMockExtensionHost(),
+			promptState: createMockPromptState(),
+			deltaTracker: createMockDeltaTracker(),
+			commandStreamManager: createMockCommandStreamManager(),
+			toolContentStreamManager: createMockToolContentStreamManager(),
+			toolHandlerRegistry: new ToolHandlerRegistry(),
+			sendUpdate: mockSendUpdate,
+			approveAction: vi.fn(),
+			respondWithText: vi.fn(),
+			sendToExtension: vi.fn(),
+			workspacePath: "/test/workspace",
+			initialModeId: "code",
+			isCancelling: vi.fn().mockReturnValue(false),
+		}
+
+		eventHandler = createSessionEventHandler(deps)
+		eventHandler.setupEventHandlers()
+	})
+
+	describe("todo list message detection", () => {
+		it("detects and sends plan update for updateTodoList tool ask message", () => {
+			const todoMessage: ClineMessage = {
+				type: "ask",
+				ask: "tool",
+				ts: Date.now(),
+				text: JSON.stringify({
+					tool: "updateTodoList",
+					todos: [
+						{ id: "1", content: "First task", status: "completed" },
+						{ id: "2", content: "Second task", status: "in_progress" },
+						{ id: "3", content: "Third task", status: "pending" },
+					],
+				}),
+			}
+
+			// Emit the message through the mock client
+			mockClient.emit("message", todoMessage)
+
+			// Verify plan update was sent
+			expect(mockSendUpdate).toHaveBeenCalledWith(
+				expect.objectContaining({
+					sessionUpdate: "plan",
+					entries: expect.arrayContaining([
+						expect.objectContaining({
+							content: "First task",
+							status: "completed",
+							priority: expect.any(String),
+						}),
+						expect.objectContaining({
+							content: "Second task",
+							status: "in_progress",
+							priority: "high", // in_progress gets high priority
+						}),
+						expect.objectContaining({
+							content: "Third task",
+							status: "pending",
+							priority: expect.any(String),
+						}),
+					]),
+				}),
+			)
+		})
+
+		it("detects and sends plan update for user_edit_todos say message", () => {
+			const editMessage: ClineMessage = {
+				type: "say",
+				say: "user_edit_todos",
+				ts: Date.now(),
+				text: JSON.stringify({
+					tool: "updateTodoList",
+					todos: [{ id: "1", content: "Edited task", status: "completed" }],
+				}),
+			}
+
+			mockClient.emit("message", editMessage)
+
+			expect(mockSendUpdate).toHaveBeenCalledWith(
+				expect.objectContaining({
+					sessionUpdate: "plan",
+					entries: [
+						expect.objectContaining({
+							content: "Edited task",
+							status: "completed",
+						}),
+					],
+				}),
+			)
+		})
+
+		it("does not send plan update for other tool ask messages", () => {
+			const otherToolMessage: ClineMessage = {
+				type: "ask",
+				ask: "tool",
+				ts: Date.now(),
+				text: JSON.stringify({
+					tool: "read_file",
+					path: "/some/file.txt",
+				}),
+			}
+
+			mockClient.emit("message", otherToolMessage)
+
+			// Should not have sent a plan update (but may send other updates)
+			const planUpdateCalls = mockSendUpdate.mock.calls.filter(
+				(call) => (call[0] as { sessionUpdate?: string })?.sessionUpdate === "plan",
+			)
+			expect(planUpdateCalls).toHaveLength(0)
+		})
+
+		it("does not send plan update for empty todo list", () => {
+			const emptyTodoMessage: ClineMessage = {
+				type: "ask",
+				ask: "tool",
+				ts: Date.now(),
+				text: JSON.stringify({
+					tool: "updateTodoList",
+					todos: [],
+				}),
+			}
+
+			mockClient.emit("message", emptyTodoMessage)
+
+			// Should not have sent a plan update for empty list
+			const planUpdateCalls = mockSendUpdate.mock.calls.filter(
+				(call) => (call[0] as { sessionUpdate?: string })?.sessionUpdate === "plan",
+			)
+			expect(planUpdateCalls).toHaveLength(0)
+		})
+	})
+
+	describe("priority assignment", () => {
+		it("assigns high priority to in_progress items", () => {
+			const todoMessage: ClineMessage = {
+				type: "ask",
+				ask: "tool",
+				ts: Date.now(),
+				text: JSON.stringify({
+					tool: "updateTodoList",
+					todos: [
+						{ id: "1", content: "Pending task", status: "pending" },
+						{ id: "2", content: "In progress task", status: "in_progress" },
+						{ id: "3", content: "Completed task", status: "completed" },
+					],
+				}),
+			}
+
+			mockClient.emit("message", todoMessage)
+
+			const planUpdateCall = mockSendUpdate.mock.calls.find(
+				(call) => (call[0] as { sessionUpdate?: string })?.sessionUpdate === "plan",
+			)
+			expect(planUpdateCall).toBeDefined()
+
+			const entries = (planUpdateCall![0] as { entries: Array<{ content: string; priority: string }> }).entries
+			const inProgressEntry = entries.find((e) => e.content === "In progress task")
+
+			expect(inProgressEntry?.priority).toBe("high")
+		})
+
+		it("assigns medium priority to pending and completed items by default", () => {
+			const todoMessage: ClineMessage = {
+				type: "ask",
+				ask: "tool",
+				ts: Date.now(),
+				text: JSON.stringify({
+					tool: "updateTodoList",
+					todos: [
+						{ id: "1", content: "Pending task", status: "pending" },
+						{ id: "2", content: "Completed task", status: "completed" },
+					],
+				}),
+			}
+
+			mockClient.emit("message", todoMessage)
+
+			const planUpdateCall = mockSendUpdate.mock.calls.find(
+				(call) => (call[0] as { sessionUpdate?: string })?.sessionUpdate === "plan",
+			)
+			expect(planUpdateCall).toBeDefined()
+
+			const entries = (planUpdateCall![0] as { entries: Array<{ content: string; priority: string }> }).entries
+			const pendingEntry = entries.find((e) => e.content === "Pending task")
+			const completedEntry = entries.find((e) => e.content === "Completed task")
+
+			expect(pendingEntry?.priority).toBe("medium")
+			expect(completedEntry?.priority).toBe("medium")
+		})
+	})
+
+	describe("message updates (streaming)", () => {
+		it("sends plan update when message is updated", () => {
+			const todoMessage: ClineMessage = {
+				type: "ask",
+				ask: "tool",
+				ts: Date.now(),
+				text: JSON.stringify({
+					tool: "updateTodoList",
+					todos: [{ id: "1", content: "Initial task", status: "pending" }],
+				}),
+			}
+
+			// First message
+			mockClient.emit("message", todoMessage)
+
+			// Updated message with more todos
+			const updatedMessage: ClineMessage = {
+				...todoMessage,
+				text: JSON.stringify({
+					tool: "updateTodoList",
+					todos: [
+						{ id: "1", content: "Initial task", status: "completed" },
+						{ id: "2", content: "New task", status: "pending" },
+					],
+				}),
+			}
+
+			mockClient.emit("messageUpdated", updatedMessage)
+
+			// Should have sent 2 plan updates (one for each message)
+			const planUpdateCalls = mockSendUpdate.mock.calls.filter(
+				(call) => (call[0] as { sessionUpdate?: string })?.sessionUpdate === "plan",
+			)
+			expect(planUpdateCalls.length).toBeGreaterThanOrEqual(2)
+		})
+	})
+
+	describe("logging", () => {
+		it("sends plan updates without verbose logging", () => {
+			const todoMessage: ClineMessage = {
+				type: "ask",
+				ask: "tool",
+				ts: Date.now(),
+				text: JSON.stringify({
+					tool: "updateTodoList",
+					todos: [{ id: "1", content: "Test task", status: "pending" }],
+				}),
+			}
+
+			mockClient.emit("message", todoMessage)
+
+			// Plan update should be sent without verbose logging
+			const planUpdateCalls = mockSendUpdate.mock.calls.filter(
+				(call) => (call[0] as { sessionUpdate?: string })?.sessionUpdate === "plan",
+			)
+			expect(planUpdateCalls).toHaveLength(1)
+		})
+	})
+
+	describe("reset behavior", () => {
+		it("continues to detect plan updates after reset", () => {
+			const todoMessage: ClineMessage = {
+				type: "ask",
+				ask: "tool",
+				ts: Date.now(),
+				text: JSON.stringify({
+					tool: "updateTodoList",
+					todos: [{ id: "1", content: "Task 1", status: "pending" }],
+				}),
+			}
+
+			mockClient.emit("message", todoMessage)
+			mockSendUpdate.mockClear()
+
+			// Reset the event handler
+			eventHandler.reset()
+
+			// Send another todo message
+			const anotherMessage: ClineMessage = {
+				...todoMessage,
+				ts: Date.now() + 1,
+				text: JSON.stringify({
+					tool: "updateTodoList",
+					todos: [{ id: "2", content: "Task 2", status: "pending" }],
+				}),
+			}
+
+			mockClient.emit("message", anotherMessage)
+
+			// Should still detect and send plan update
+			const planUpdateCalls = mockSendUpdate.mock.calls.filter(
+				(call) => (call[0] as { sessionUpdate?: string })?.sessionUpdate === "plan",
+			)
+			expect(planUpdateCalls).toHaveLength(1)
+		})
+	})
+})

apps/cli/src/acp/__tests__/session.test.ts

@@ -0,0 +1,301 @@
+import type * as acp from "@agentclientprotocol/sdk"
+import { AgentLoopState } from "@/agent/agent-state.js"
+
+// Track registered event handlers for simulation
+type EventHandler = (data: unknown) => void
+const clientEventHandlers: Map<string, EventHandler[]> = new Map()
+
+vi.mock("@/agent/extension-host.js", () => {
+	const mockClient = {
+		on: vi.fn().mockImplementation((event: string, handler: EventHandler) => {
+			const handlers = clientEventHandlers.get(event) || []
+			handlers.push(handler)
+			clientEventHandlers.set(event, handlers)
+			return mockClient
+		}),
+		off: vi.fn().mockReturnThis(),
+		respond: vi.fn(),
+		approve: vi.fn(),
+		reject: vi.fn(),
+		getAgentState: vi.fn().mockReturnValue({
+			state: AgentLoopState.RUNNING,
+			isRunning: true,
+			isStreaming: false,
+			currentAsk: null,
+		}),
+	}
+
+	return {
+		ExtensionHost: vi.fn().mockImplementation(() => ({
+			client: mockClient,
+			activate: vi.fn().mockResolvedValue(undefined),
+			dispose: vi.fn().mockResolvedValue(undefined),
+			sendToExtension: vi.fn(),
+			// Add on/off methods for extension host events (e.g., extensionWebviewMessage)
+			on: vi.fn().mockReturnThis(),
+			off: vi.fn().mockReturnThis(),
+		})),
+	}
+})
+
+/**
+ * Simulate the extension responding to a cancel by emitting a state change to a terminal state.
+ */
+function simulateExtensionCancelResponse(): void {
+	const handlers = clientEventHandlers.get("stateChange") || []
+	handlers.forEach((handler) => {
+		handler({
+			previousState: { state: AgentLoopState.RUNNING, isRunning: true, isStreaming: false },
+			currentState: { state: AgentLoopState.IDLE, isRunning: false, isStreaming: false },
+		})
+	})
+}
+
+import { AcpSession, type AcpSessionOptions } from "../session.js"
+import { ExtensionHost } from "@/agent/extension-host.js"
+
+describe("AcpSession", () => {
+	let mockConnection: acp.AgentSideConnection
+
+	const defaultOptions: AcpSessionOptions = {
+		extensionPath: "/test/extension",
+		provider: "openrouter",
+		apiKey: "test-api-key",
+		model: "test-model",
+		mode: "code",
+	}
+
+	beforeEach(() => {
+		// Clear registered event handlers between tests
+		clientEventHandlers.clear()
+
+		mockConnection = {
+			sessionUpdate: vi.fn().mockResolvedValue(undefined),
+			requestPermission: vi.fn().mockResolvedValue({
+				outcome: { outcome: "selected", optionId: "allow" },
+			}),
+			readTextFile: vi.fn().mockResolvedValue({ content: "test content" }),
+			writeTextFile: vi.fn().mockResolvedValue({}),
+			createTerminal: vi.fn(),
+			extMethod: vi.fn(),
+			extNotification: vi.fn(),
+			signal: new AbortController().signal,
+			closed: Promise.resolve(),
+		} as unknown as acp.AgentSideConnection
+
+		vi.clearAllMocks()
+	})
+
+	afterEach(() => {
+		vi.clearAllMocks()
+		clientEventHandlers.clear()
+	})
+
+	describe("create", () => {
+		it("should create a session with a unique ID", async () => {
+			const session = await AcpSession.create({
+				sessionId: "test-session-1",
+				cwd: "/test/workspace",
+				connection: mockConnection,
+				options: defaultOptions,
+				deps: {},
+			})
+
+			expect(session).toBeDefined()
+			expect(session.getSessionId()).toBe("test-session-1")
+		})
+
+		it("should create ExtensionHost with correct config", async () => {
+			await AcpSession.create({
+				sessionId: "test-session-2",
+				cwd: "/test/workspace",
+				connection: mockConnection,
+				options: defaultOptions,
+				deps: {},
+			})
+
+			expect(ExtensionHost).toHaveBeenCalledWith(
+				expect.objectContaining({
+					extensionPath: "/test/extension",
+					workspacePath: "/test/workspace",
+					provider: "openrouter",
+					apiKey: "test-api-key",
+					model: "test-model",
+					mode: "code",
+				}),
+			)
+		})
+
+		it("should accept client capabilities", async () => {
+			const session = await AcpSession.create({
+				sessionId: "test-session-3",
+				cwd: "/test/workspace",
+				connection: mockConnection,
+				options: defaultOptions,
+				deps: {},
+			})
+
+			expect(session).toBeDefined()
+		})
+
+		it("should activate the extension host", async () => {
+			await AcpSession.create({
+				sessionId: "test-session-4",
+				cwd: "/test/workspace",
+				connection: mockConnection,
+				options: defaultOptions,
+				deps: {},
+			})
+
+			const mockHostInstance = vi.mocked(ExtensionHost).mock.results[0]!.value
+			expect(mockHostInstance.activate).toHaveBeenCalled()
+		})
+	})
+
+	describe("prompt", () => {
+		it("should send a task to the extension host", async () => {
+			const session = await AcpSession.create({
+				sessionId: "test-session",
+				cwd: "/test/workspace",
+				connection: mockConnection,
+				options: defaultOptions,
+				deps: {},
+			})
+
+			const mockHostInstance = vi.mocked(ExtensionHost).mock.results[0]!.value
+
+			// Start the prompt (don't await - it waits for taskCompleted event)
+			const promptPromise = session.prompt({
+				sessionId: "test-session",
+				prompt: [{ type: "text", text: "Hello, world!" }],
+			})
+
+			// Verify the task was sent
+			expect(mockHostInstance.sendToExtension).toHaveBeenCalledWith(
+				expect.objectContaining({
+					type: "newTask",
+					text: "Hello, world!",
+				}),
+			)
+
+			// Cancel to resolve the promise - simulate extension responding to cancel
+			session.cancel()
+			simulateExtensionCancelResponse()
+			const result = await promptPromise
+			expect(result.stopReason).toBe("cancelled")
+		})
+
+		it("should handle image prompts", async () => {
+			const session = await AcpSession.create({
+				sessionId: "test-session",
+				cwd: "/test/workspace",
+				connection: mockConnection,
+				options: defaultOptions,
+				deps: {},
+			})
+
+			const mockHostInstance = vi.mocked(ExtensionHost).mock.results[0]!.value
+
+			const promptPromise = session.prompt({
+				sessionId: "test-session",
+				prompt: [
+					{ type: "text", text: "Describe this image" },
+					{ type: "image", mimeType: "image/png", data: "base64data" },
+				],
+			})
+
+			// Images are extracted as raw base64 data, text includes [image content] placeholder
+			expect(mockHostInstance.sendToExtension).toHaveBeenCalledWith(
+				expect.objectContaining({
+					type: "newTask",
+					images: expect.arrayContaining(["base64data"]),
+				}),
+			)
+
+			session.cancel()
+			simulateExtensionCancelResponse()
+			await promptPromise
+		})
+	})
+
+	describe("cancel", () => {
+		it("should send cancel message to extension host", async () => {
+			const session = await AcpSession.create({
+				sessionId: "test-session",
+				cwd: "/test/workspace",
+				connection: mockConnection,
+				options: defaultOptions,
+				deps: {},
+			})
+
+			const mockHostInstance = vi.mocked(ExtensionHost).mock.results[0]!.value
+
+			// Start a prompt first
+			const promptPromise = session.prompt({
+				sessionId: "test-session",
+				prompt: [{ type: "text", text: "Hello" }],
+			})
+
+			// Cancel and simulate extension responding
+			session.cancel()
+			simulateExtensionCancelResponse()
+
+			expect(mockHostInstance.sendToExtension).toHaveBeenCalledWith({ type: "cancelTask" })
+
+			await promptPromise
+		})
+	})
+
+	describe("setMode", () => {
+		it("should update the session mode", async () => {
+			const session = await AcpSession.create({
+				sessionId: "test-session",
+				cwd: "/test/workspace",
+				connection: mockConnection,
+				options: defaultOptions,
+				deps: {},
+			})
+
+			const mockHostInstance = vi.mocked(ExtensionHost).mock.results[0]!.value
+
+			session.setMode("architect")
+
+			expect(mockHostInstance.sendToExtension).toHaveBeenCalledWith({
+				type: "updateSettings",
+				updatedSettings: { mode: "architect" },
+			})
+		})
+	})
+
+	describe("dispose", () => {
+		it("should dispose the extension host", async () => {
+			const session = await AcpSession.create({
+				sessionId: "test-session",
+				cwd: "/test/workspace",
+				connection: mockConnection,
+				options: defaultOptions,
+				deps: {},
+			})
+
+			const mockHostInstance = vi.mocked(ExtensionHost).mock.results[0]!.value
+
+			await session.dispose()
+
+			expect(mockHostInstance.dispose).toHaveBeenCalled()
+		})
+	})
+
+	describe("getSessionId", () => {
+		it("should return the session ID", async () => {
+			const session = await AcpSession.create({
+				sessionId: "my-unique-session-id",
+				cwd: "/test/workspace",
+				connection: mockConnection,
+				options: defaultOptions,
+				deps: {},
+			})
+
+			expect(session.getSessionId()).toBe("my-unique-session-id")
+		})
+	})
+})

apps/cli/src/acp/__tests__/tool-content-stream.test.ts

@@ -0,0 +1,495 @@
+/**
+ * Tests for ToolContentStreamManager
+ *
+ * Tests the tool content (file creates/edits) streaming functionality
+ * extracted from session.ts.
+ */
+
+import type { ClineMessage } from "@roo-code/types"
+
+import { DeltaTracker } from "../delta-tracker.js"
+import { ToolContentStreamManager } from "../tool-content-stream.js"
+import { NullLogger } from "../interfaces.js"
+import type { SendUpdateFn } from "../interfaces.js"
+
+describe("ToolContentStreamManager", () => {
+	let deltaTracker: DeltaTracker
+	let sendUpdate: SendUpdateFn
+	let sentUpdates: Array<Record<string, unknown>>
+	let manager: ToolContentStreamManager
+
+	beforeEach(() => {
+		deltaTracker = new DeltaTracker()
+		sentUpdates = []
+		sendUpdate = (update) => {
+			sentUpdates.push(update as Record<string, unknown>)
+		}
+		manager = new ToolContentStreamManager({
+			deltaTracker,
+			sendUpdate,
+			logger: new NullLogger(),
+		})
+	})
+
+	describe("isToolAskMessage", () => {
+		it("returns true for tool ask messages", () => {
+			const message: ClineMessage = {
+				type: "ask",
+				ask: "tool",
+				ts: Date.now(),
+				text: "{}",
+			}
+			expect(manager.isToolAskMessage(message)).toBe(true)
+		})
+
+		it("returns false for other ask types", () => {
+			const message: ClineMessage = {
+				type: "ask",
+				ask: "command",
+				ts: Date.now(),
+				text: "npm test",
+			}
+			expect(manager.isToolAskMessage(message)).toBe(false)
+		})
+
+		it("returns false for say messages", () => {
+			const message: ClineMessage = {
+				type: "say",
+				say: "text",
+				ts: Date.now(),
+				text: "hello",
+			}
+			expect(manager.isToolAskMessage(message)).toBe(false)
+		})
+	})
+
+	describe("handleToolContentStreaming", () => {
+		describe("file write tool detection", () => {
+			const fileWriteTools = [
+				"newFileCreated",
+				"write_to_file",
+				"create_file",
+				"editedExistingFile",
+				"apply_diff",
+				"modify_file",
+			]
+
+			fileWriteTools.forEach((toolName) => {
+				it(`handles ${toolName} as a file write tool`, () => {
+					const message: ClineMessage = {
+						type: "ask",
+						ask: "tool",
+						ts: 12345,
+						text: JSON.stringify({
+							tool: toolName,
+							path: "test.ts",
+							content: "content",
+						}),
+						partial: true,
+					}
+
+					const result = manager.handleToolContentStreaming(message)
+					expect(result).toBe(true)
+					// Should send header since it's a file write tool
+					expect(sentUpdates.length).toBeGreaterThan(0)
+				})
+			})
+
+			it("skips non-file tools", () => {
+				const message: ClineMessage = {
+					type: "ask",
+					ask: "tool",
+					ts: 12345,
+					text: JSON.stringify({
+						tool: "read_file",
+						path: "test.ts",
+					}),
+					partial: true,
+				}
+
+				const result = manager.handleToolContentStreaming(message)
+				expect(result).toBe(true) // Handled by skipping
+				expect(sentUpdates.length).toBe(0) // Nothing sent
+			})
+		})
+
+		describe("header management", () => {
+			it("sends header on first valid path", () => {
+				const message: ClineMessage = {
+					type: "ask",
+					ask: "tool",
+					ts: 12345,
+					text: JSON.stringify({
+						tool: "write_to_file",
+						path: "src/index.ts",
+						content: "",
+					}),
+					partial: true,
+				}
+
+				manager.handleToolContentStreaming(message)
+
+				expect(sentUpdates.length).toBe(1)
+				expect(sentUpdates[0]).toEqual({
+					sessionUpdate: "agent_message_chunk",
+					content: { type: "text", text: "\n**Creating src/index.ts**\n```\n" },
+				})
+			})
+
+			it("only sends header once per message", () => {
+				const ts = 12345
+
+				// First call
+				manager.handleToolContentStreaming({
+					type: "ask",
+					ask: "tool",
+					ts,
+					text: JSON.stringify({
+						tool: "write_to_file",
+						path: "test.ts",
+						content: "line 1",
+					}),
+					partial: true,
+				})
+
+				const headerCount1 = sentUpdates.filter((u) =>
+					((u.content as { text: string }).text || "").includes("**Creating"),
+				).length
+
+				// Second call with same ts
+				manager.handleToolContentStreaming({
+					type: "ask",
+					ask: "tool",
+					ts,
+					text: JSON.stringify({
+						tool: "write_to_file",
+						path: "test.ts",
+						content: "line 1\nline 2",
+					}),
+					partial: true,
+				})
+
+				const headerCount2 = sentUpdates.filter((u) =>
+					((u.content as { text: string }).text || "").includes("**Creating"),
+				).length
+
+				expect(headerCount1).toBe(1)
+				expect(headerCount2).toBe(1) // Still 1, no duplicate
+			})
+
+			it("waits for valid path before sending header", () => {
+				// Path without extension is not valid
+				const message: ClineMessage = {
+					type: "ask",
+					ask: "tool",
+					ts: 12345,
+					text: JSON.stringify({
+						tool: "write_to_file",
+						path: "incomplete",
+						content: "content",
+					}),
+					partial: true,
+				}
+
+				manager.handleToolContentStreaming(message)
+				expect(sentUpdates.length).toBe(0) // No header yet
+			})
+
+			it("validates path has file extension", () => {
+				const validPaths = ["test.ts", "README.md", "config.json", "src/utils.js"]
+				const invalidPaths = ["test", "src/folder/", "noextension"]
+
+				validPaths.forEach((path) => {
+					sentUpdates.length = 0
+					manager = new ToolContentStreamManager({
+						deltaTracker: new DeltaTracker(),
+						sendUpdate,
+						logger: new NullLogger(),
+					})
+
+					manager.handleToolContentStreaming({
+						type: "ask",
+						ask: "tool",
+						ts: Date.now(),
+						text: JSON.stringify({ tool: "write_to_file", path, content: "" }),
+						partial: true,
+					})
+
+					expect(sentUpdates.length).toBeGreaterThan(0)
+				})
+
+				invalidPaths.forEach((path) => {
+					sentUpdates.length = 0
+					manager = new ToolContentStreamManager({
+						deltaTracker: new DeltaTracker(),
+						sendUpdate,
+						logger: new NullLogger(),
+					})
+
+					manager.handleToolContentStreaming({
+						type: "ask",
+						ask: "tool",
+						ts: Date.now(),
+						text: JSON.stringify({ tool: "write_to_file", path, content: "x" }),
+						partial: true,
+					})
+
+					expect(sentUpdates.length).toBe(0)
+				})
+			})
+		})
+
+		describe("content streaming", () => {
+			it("streams content deltas", () => {
+				const ts = 12345
+
+				// First chunk
+				manager.handleToolContentStreaming({
+					type: "ask",
+					ask: "tool",
+					ts,
+					text: JSON.stringify({
+						tool: "write_to_file",
+						path: "test.ts",
+						content: "const x = 1;",
+					}),
+					partial: true,
+				})
+
+				// Header + content
+				expect(sentUpdates.length).toBe(2)
+				expect(sentUpdates[1]).toEqual({
+					sessionUpdate: "agent_message_chunk",
+					content: { type: "text", text: "const x = 1;" },
+				})
+
+				// Second chunk with more content
+				manager.handleToolContentStreaming({
+					type: "ask",
+					ask: "tool",
+					ts,
+					text: JSON.stringify({
+						tool: "write_to_file",
+						path: "test.ts",
+						content: "const x = 1;\nconst y = 2;",
+					}),
+					partial: true,
+				})
+
+				// Should only send the delta
+				expect(sentUpdates.length).toBe(3)
+				expect(sentUpdates[2]).toEqual({
+					sessionUpdate: "agent_message_chunk",
+					content: { type: "text", text: "\nconst y = 2;" },
+				})
+			})
+
+			it("handles multiple tool streams independently", () => {
+				// First tool
+				manager.handleToolContentStreaming({
+					type: "ask",
+					ask: "tool",
+					ts: 1000,
+					text: JSON.stringify({
+						tool: "write_to_file",
+						path: "file1.ts",
+						content: "content1",
+					}),
+					partial: true,
+				})
+
+				// Second tool
+				manager.handleToolContentStreaming({
+					type: "ask",
+					ask: "tool",
+					ts: 2000,
+					text: JSON.stringify({
+						tool: "write_to_file",
+						path: "file2.ts",
+						content: "content2",
+					}),
+					partial: true,
+				})
+
+				// Both should get headers
+				const headers = sentUpdates.filter((u) =>
+					((u.content as { text: string }).text || "").includes("**Creating"),
+				)
+				expect(headers.length).toBe(2)
+			})
+		})
+
+		describe("completion", () => {
+			it("sends closing code fence on complete", () => {
+				const ts = 12345
+
+				// Partial message
+				manager.handleToolContentStreaming({
+					type: "ask",
+					ask: "tool",
+					ts,
+					text: JSON.stringify({
+						tool: "write_to_file",
+						path: "test.ts",
+						content: "content",
+					}),
+					partial: true,
+				})
+
+				sentUpdates.length = 0 // Clear
+
+				// Complete message
+				manager.handleToolContentStreaming({
+					type: "ask",
+					ask: "tool",
+					ts,
+					text: JSON.stringify({
+						tool: "write_to_file",
+						path: "test.ts",
+						content: "content",
+					}),
+					partial: false,
+				})
+
+				expect(sentUpdates[0]).toEqual({
+					sessionUpdate: "agent_message_chunk",
+					content: { type: "text", text: "\n```\n" },
+				})
+			})
+
+			it("cleans up header tracking on complete", () => {
+				const ts = 12345
+
+				// Partial
+				manager.handleToolContentStreaming({
+					type: "ask",
+					ask: "tool",
+					ts,
+					text: JSON.stringify({
+						tool: "write_to_file",
+						path: "test.ts",
+						content: "content",
+					}),
+					partial: true,
+				})
+
+				expect(manager.getActiveHeaderCount()).toBe(1)
+
+				// Complete
+				manager.handleToolContentStreaming({
+					type: "ask",
+					ask: "tool",
+					ts,
+					text: JSON.stringify({
+						tool: "write_to_file",
+						path: "test.ts",
+						content: "content",
+					}),
+					partial: false,
+				})
+
+				expect(manager.getActiveHeaderCount()).toBe(0)
+			})
+
+			it("does not send code fence if no header was sent", () => {
+				const ts = 12345
+
+				// Complete message without prior partial (no header sent)
+				manager.handleToolContentStreaming({
+					type: "ask",
+					ask: "tool",
+					ts,
+					text: JSON.stringify({
+						tool: "write_to_file",
+						path: "test.ts",
+						content: "content",
+					}),
+					partial: false,
+				})
+
+				// Should not send closing fence
+				const closingFences = sentUpdates.filter((u) =>
+					((u.content as { text: string }).text || "").includes("```"),
+				)
+				expect(closingFences.length).toBe(0)
+			})
+		})
+
+		describe("JSON parsing", () => {
+			it("handles invalid JSON gracefully", () => {
+				const message: ClineMessage = {
+					type: "ask",
+					ask: "tool",
+					ts: 12345,
+					text: "{incomplete json",
+					partial: true,
+				}
+
+				const result = manager.handleToolContentStreaming(message)
+				expect(result).toBe(true) // Handled by returning early
+				expect(sentUpdates.length).toBe(0)
+			})
+
+			it("handles empty text", () => {
+				const message: ClineMessage = {
+					type: "ask",
+					ask: "tool",
+					ts: 12345,
+					text: "",
+					partial: true,
+				}
+
+				const result = manager.handleToolContentStreaming(message)
+				expect(result).toBe(true)
+				expect(sentUpdates.length).toBe(0)
+			})
+		})
+	})
+
+	describe("reset", () => {
+		it("clears header tracking", () => {
+			manager.handleToolContentStreaming({
+				type: "ask",
+				ask: "tool",
+				ts: 12345,
+				text: JSON.stringify({
+					tool: "write_to_file",
+					path: "test.ts",
+					content: "content",
+				}),
+				partial: true,
+			})
+
+			expect(manager.getActiveHeaderCount()).toBe(1)
+
+			manager.reset()
+			expect(manager.getActiveHeaderCount()).toBe(0)
+		})
+	})
+
+	describe("getActiveHeaderCount", () => {
+		it("returns 0 initially", () => {
+			expect(manager.getActiveHeaderCount()).toBe(0)
+		})
+
+		it("returns correct count after streaming", () => {
+			manager.handleToolContentStreaming({
+				type: "ask",
+				ask: "tool",
+				ts: 1000,
+				text: JSON.stringify({ tool: "write_to_file", path: "a.ts", content: "" }),
+				partial: true,
+			})
+
+			manager.handleToolContentStreaming({
+				type: "ask",
+				ask: "tool",
+				ts: 2000,
+				text: JSON.stringify({ tool: "write_to_file", path: "b.ts", content: "" }),
+				partial: true,
+			})
+
+			expect(manager.getActiveHeaderCount()).toBe(2)
+		})
+	})
+})

apps/cli/src/acp/__tests__/tool-handler.test.ts

@@ -0,0 +1,495 @@
+/**
+ * Tool Handler Unit Tests
+ *
+ * Tests for the ToolHandler abstraction and ToolHandlerRegistry.
+ */
+
+import type { ClineMessage, ClineAsk } from "@roo-code/types"
+
+import {
+	ToolHandlerRegistry,
+	CommandToolHandler,
+	FileEditToolHandler,
+	FileReadToolHandler,
+	SearchToolHandler,
+	ListFilesToolHandler,
+	DefaultToolHandler,
+	type ToolHandlerContext,
+} from "../tool-handler.js"
+import { parseToolFromMessage } from "../translator.js"
+import { NullLogger } from "../interfaces.js"
+
+// =============================================================================
+// Test Utilities
+// =============================================================================
+
+const testLogger = new NullLogger()
+
+function createContext(message: ClineMessage, ask: ClineAsk, workspacePath = "/workspace"): ToolHandlerContext {
+	return {
+		message,
+		ask,
+		workspacePath,
+		toolInfo: parseToolFromMessage(message, workspacePath),
+		logger: testLogger,
+	}
+}
+
+function createToolMessage(tool: string, params: Record<string, unknown> = {}): ClineMessage {
+	return {
+		ts: Date.now(),
+		type: "say",
+		say: "text",
+		text: JSON.stringify({ tool, ...params }),
+	}
+}
+
+// =============================================================================
+// CommandToolHandler Tests
+// =============================================================================
+
+describe("CommandToolHandler", () => {
+	const handler = new CommandToolHandler()
+
+	describe("canHandle", () => {
+		it("should handle command asks", () => {
+			const context = createContext(createToolMessage("execute_command", { command: "ls" }), "command")
+			expect(handler.canHandle(context)).toBe(true)
+		})
+
+		it("should not handle tool asks", () => {
+			const context = createContext(createToolMessage("read_file", { path: "test.ts" }), "tool")
+			expect(handler.canHandle(context)).toBe(false)
+		})
+
+		it("should not handle browser_action_launch asks", () => {
+			const context = createContext(createToolMessage("browser_action", {}), "browser_action_launch")
+			expect(handler.canHandle(context)).toBe(false)
+		})
+	})
+
+	describe("handle", () => {
+		it("should return execute kind for commands", () => {
+			const context = createContext(createToolMessage("execute_command", { command: "npm test" }), "command")
+			const result = handler.handle(context)
+
+			expect(result.initialUpdate).toMatchObject({
+				sessionUpdate: "tool_call",
+				kind: "execute",
+				status: "in_progress",
+			})
+		})
+
+		it("should track as pending command", () => {
+			const message = createToolMessage("execute_command", { command: "npm test" })
+			const context = createContext(message, "command")
+			const result = handler.handle(context)
+
+			expect(result.trackAsPendingCommand).toBeDefined()
+			expect(result.trackAsPendingCommand?.command).toBe(message.text)
+			expect(result.trackAsPendingCommand?.ts).toBe(message.ts)
+		})
+
+		it("should not include completion update", () => {
+			const context = createContext(createToolMessage("execute_command", { command: "ls" }), "command")
+			const result = handler.handle(context)
+
+			expect(result.completionUpdate).toBeUndefined()
+		})
+	})
+})
+
+// =============================================================================
+// FileEditToolHandler Tests
+// =============================================================================
+
+describe("FileEditToolHandler", () => {
+	const handler = new FileEditToolHandler()
+
+	describe("canHandle", () => {
+		it("should handle write_to_file tool", () => {
+			const context = createContext(createToolMessage("write_to_file", { path: "test.ts" }), "tool")
+			expect(handler.canHandle(context)).toBe(true)
+		})
+
+		it("should handle apply_diff tool", () => {
+			const context = createContext(createToolMessage("apply_diff", { path: "test.ts" }), "tool")
+			expect(handler.canHandle(context)).toBe(true)
+		})
+
+		it("should handle create_file tool", () => {
+			const context = createContext(createToolMessage("create_file", { path: "new.ts" }), "tool")
+			expect(handler.canHandle(context)).toBe(true)
+		})
+
+		it("should handle newFileCreated tool", () => {
+			const context = createContext(createToolMessage("newFileCreated", { path: "new.ts" }), "tool")
+			expect(handler.canHandle(context)).toBe(true)
+		})
+
+		it("should handle editedExistingFile tool", () => {
+			const context = createContext(createToolMessage("editedExistingFile", { path: "test.ts" }), "tool")
+			expect(handler.canHandle(context)).toBe(true)
+		})
+
+		it("should not handle read_file tool", () => {
+			const context = createContext(createToolMessage("read_file", { path: "test.ts" }), "tool")
+			expect(handler.canHandle(context)).toBe(false)
+		})
+
+		it("should not handle command asks", () => {
+			const context = createContext(createToolMessage("write_to_file", { path: "test.ts" }), "command")
+			expect(handler.canHandle(context)).toBe(false)
+		})
+	})
+
+	describe("handle", () => {
+		it("should return edit kind", () => {
+			const context = createContext(createToolMessage("write_to_file", { path: "test.ts" }), "tool")
+			const result = handler.handle(context)
+
+			expect(result.initialUpdate).toMatchObject({
+				sessionUpdate: "tool_call",
+				kind: "edit",
+				status: "in_progress",
+			})
+		})
+
+		it("should include completion update", () => {
+			const context = createContext(createToolMessage("write_to_file", { path: "test.ts" }), "tool")
+			const result = handler.handle(context)
+
+			expect(result.completionUpdate).toMatchObject({
+				sessionUpdate: "tool_call_update",
+				status: "completed",
+			})
+		})
+
+		it("should not track as pending command", () => {
+			const context = createContext(createToolMessage("write_to_file", { path: "test.ts" }), "tool")
+			const result = handler.handle(context)
+
+			expect(result.trackAsPendingCommand).toBeUndefined()
+		})
+	})
+})
+
+// =============================================================================
+// FileReadToolHandler Tests
+// =============================================================================
+
+describe("FileReadToolHandler", () => {
+	const handler = new FileReadToolHandler()
+
+	describe("canHandle", () => {
+		it("should handle read_file tool", () => {
+			const context = createContext(createToolMessage("read_file", { path: "test.ts" }), "tool")
+			expect(handler.canHandle(context)).toBe(true)
+		})
+
+		it("should handle readFile tool", () => {
+			const context = createContext(createToolMessage("readFile", { path: "test.ts" }), "tool")
+			expect(handler.canHandle(context)).toBe(true)
+		})
+
+		it("should not handle write_to_file tool", () => {
+			const context = createContext(createToolMessage("write_to_file", { path: "test.ts" }), "tool")
+			expect(handler.canHandle(context)).toBe(false)
+		})
+
+		it("should not handle command asks", () => {
+			const context = createContext(createToolMessage("read_file", { path: "test.ts" }), "command")
+			expect(handler.canHandle(context)).toBe(false)
+		})
+	})
+
+	describe("handle", () => {
+		it("should return read kind", () => {
+			const context = createContext(createToolMessage("read_file", { path: "test.ts" }), "tool")
+			const result = handler.handle(context)
+
+			expect(result.initialUpdate).toMatchObject({
+				sessionUpdate: "tool_call",
+				kind: "read",
+				status: "in_progress",
+			})
+		})
+
+		it("should include completion update", () => {
+			const context = createContext(createToolMessage("read_file", { path: "test.ts" }), "tool")
+			const result = handler.handle(context)
+
+			expect(result.completionUpdate).toMatchObject({
+				sessionUpdate: "tool_call_update",
+				status: "completed",
+			})
+		})
+	})
+})
+
+// =============================================================================
+// SearchToolHandler Tests
+// =============================================================================
+
+describe("SearchToolHandler", () => {
+	const handler = new SearchToolHandler()
+
+	describe("canHandle", () => {
+		it("should handle search_files tool", () => {
+			const context = createContext(createToolMessage("search_files", { regex: "test" }), "tool")
+			expect(handler.canHandle(context)).toBe(true)
+		})
+
+		it("should handle searchFiles tool", () => {
+			const context = createContext(createToolMessage("searchFiles", { regex: "test" }), "tool")
+			expect(handler.canHandle(context)).toBe(true)
+		})
+
+		it("should handle codebase_search tool", () => {
+			const context = createContext(createToolMessage("codebase_search", { query: "test" }), "tool")
+			expect(handler.canHandle(context)).toBe(true)
+		})
+
+		it("should handle grep tool", () => {
+			const context = createContext(createToolMessage("grep", { pattern: "test" }), "tool")
+			expect(handler.canHandle(context)).toBe(true)
+		})
+
+		it("should not handle custom tool with search in name (exact matching)", () => {
+			const context = createContext(createToolMessage("custom_search_tool", {}), "tool")
+			// With exact matching, "custom_search_tool" won't match the search category
+			expect(handler.canHandle(context)).toBe(false)
+		})
+
+		it("should not handle read_file tool", () => {
+			const context = createContext(createToolMessage("read_file", { path: "test.ts" }), "tool")
+			expect(handler.canHandle(context)).toBe(false)
+		})
+	})
+
+	describe("handle", () => {
+		it("should return search kind", () => {
+			const context = createContext(createToolMessage("search_files", { regex: "test" }), "tool")
+			const result = handler.handle(context)
+
+			expect(result.initialUpdate).toMatchObject({
+				sessionUpdate: "tool_call",
+				kind: "search",
+				status: "in_progress",
+			})
+		})
+
+		it("should format search results in completion", () => {
+			const searchResults = "Found 5 results.\n\n# src/file1.ts\n  1 | match\n\n# src/file2.ts\n  2 | match"
+			const context = createContext(createToolMessage("search_files", { content: searchResults }), "tool")
+			const result = handler.handle(context)
+
+			expect(result.completionUpdate).toMatchObject({
+				sessionUpdate: "tool_call_update",
+				status: "completed",
+			})
+
+			// Content should be formatted - cast to access content property
+			const completionUpdate = result.completionUpdate as Record<string, unknown>
+			expect(completionUpdate?.content).toBeDefined()
+		})
+	})
+})
+
+// =============================================================================
+// ListFilesToolHandler Tests
+// =============================================================================
+
+describe("ListFilesToolHandler", () => {
+	const handler = new ListFilesToolHandler()
+
+	describe("canHandle", () => {
+		it("should handle list_files tool", () => {
+			const context = createContext(createToolMessage("list_files", { path: "src" }), "tool")
+			expect(handler.canHandle(context)).toBe(true)
+		})
+
+		it("should handle listFiles tool", () => {
+			const context = createContext(createToolMessage("listFiles", { path: "src" }), "tool")
+			expect(handler.canHandle(context)).toBe(true)
+		})
+
+		it("should handle listFilesTopLevel tool", () => {
+			const context = createContext(createToolMessage("listFilesTopLevel", { path: "src" }), "tool")
+			expect(handler.canHandle(context)).toBe(true)
+		})
+
+		it("should handle listFilesRecursive tool", () => {
+			const context = createContext(createToolMessage("listFilesRecursive", { path: "src" }), "tool")
+			expect(handler.canHandle(context)).toBe(true)
+		})
+
+		it("should not handle read_file tool", () => {
+			const context = createContext(createToolMessage("read_file", { path: "test.ts" }), "tool")
+			expect(handler.canHandle(context)).toBe(false)
+		})
+	})
+
+	describe("handle", () => {
+		it("should return read kind", () => {
+			const context = createContext(createToolMessage("list_files", { path: "src" }), "tool")
+			const result = handler.handle(context)
+
+			expect(result.initialUpdate).toMatchObject({
+				sessionUpdate: "tool_call",
+				kind: "read",
+				status: "in_progress",
+			})
+		})
+	})
+})
+
+// =============================================================================
+// DefaultToolHandler Tests
+// =============================================================================
+
+describe("DefaultToolHandler", () => {
+	const handler = new DefaultToolHandler()
+
+	describe("canHandle", () => {
+		it("should always return true", () => {
+			const context1 = createContext(createToolMessage("unknown_tool", {}), "tool")
+			const context2 = createContext(createToolMessage("custom_operation", {}), "tool")
+			const context3 = createContext(createToolMessage("any_tool", {}), "browser_action_launch")
+
+			expect(handler.canHandle(context1)).toBe(true)
+			expect(handler.canHandle(context2)).toBe(true)
+			expect(handler.canHandle(context3)).toBe(true)
+		})
+	})
+
+	describe("handle", () => {
+		it("should map tool kind from tool name (exact matching)", () => {
+			// Use exact tool name from TOOL_CATEGORIES.think
+			const context = createContext(createToolMessage("think", {}), "tool")
+			const result = handler.handle(context)
+
+			expect(result.initialUpdate).toMatchObject({
+				sessionUpdate: "tool_call",
+				kind: "think",
+				status: "in_progress",
+			})
+		})
+
+		it("should return other kind for unknown tools (exact matching)", () => {
+			// Tool names that don't exactly match categories return "other"
+			const context = createContext(createToolMessage("think_about_it", {}), "tool")
+			const result = handler.handle(context)
+
+			expect(result.initialUpdate).toMatchObject({
+				sessionUpdate: "tool_call",
+				kind: "other",
+				status: "in_progress",
+			})
+		})
+
+		it("should include completion update", () => {
+			const context = createContext(createToolMessage("custom_tool", {}), "tool")
+			const result = handler.handle(context)
+
+			expect(result.completionUpdate).toMatchObject({
+				sessionUpdate: "tool_call_update",
+				status: "completed",
+			})
+		})
+	})
+})
+
+// =============================================================================
+// ToolHandlerRegistry Tests
+// =============================================================================
+
+describe("ToolHandlerRegistry", () => {
+	describe("getHandler", () => {
+		const registry = new ToolHandlerRegistry()
+
+		it("should return CommandToolHandler for command asks", () => {
+			const context = createContext(createToolMessage("execute_command", {}), "command")
+			const handler = registry.getHandler(context)
+
+			expect(handler).toBeInstanceOf(CommandToolHandler)
+		})
+
+		it("should return FileEditToolHandler for edit tools", () => {
+			const context = createContext(createToolMessage("write_to_file", { path: "test.ts" }), "tool")
+			const handler = registry.getHandler(context)
+
+			expect(handler).toBeInstanceOf(FileEditToolHandler)
+		})
+
+		it("should return FileReadToolHandler for read tools", () => {
+			const context = createContext(createToolMessage("read_file", { path: "test.ts" }), "tool")
+			const handler = registry.getHandler(context)
+
+			expect(handler).toBeInstanceOf(FileReadToolHandler)
+		})
+
+		it("should return SearchToolHandler for search tools", () => {
+			const context = createContext(createToolMessage("search_files", {}), "tool")
+			const handler = registry.getHandler(context)
+
+			expect(handler).toBeInstanceOf(SearchToolHandler)
+		})
+
+		it("should return ListFilesToolHandler for list tools", () => {
+			const context = createContext(createToolMessage("list_files", {}), "tool")
+			const handler = registry.getHandler(context)
+
+			expect(handler).toBeInstanceOf(ListFilesToolHandler)
+		})
+
+		it("should return DefaultToolHandler for unknown tools", () => {
+			const context = createContext(createToolMessage("unknown_tool", {}), "tool")
+			const handler = registry.getHandler(context)
+
+			expect(handler).toBeInstanceOf(DefaultToolHandler)
+		})
+	})
+
+	describe("handle", () => {
+		const registry = new ToolHandlerRegistry()
+
+		it("should dispatch to correct handler and return result", () => {
+			const context = createContext(createToolMessage("execute_command", {}), "command")
+			const result = registry.handle(context)
+
+			expect(result.initialUpdate).toMatchObject({
+				sessionUpdate: "tool_call",
+				kind: "execute",
+				status: "in_progress",
+			})
+			expect(result.trackAsPendingCommand).toBeDefined()
+		})
+	})
+
+	describe("createContext", () => {
+		it("should create a valid context", () => {
+			const message = createToolMessage("read_file", { path: "test.ts" })
+			const context = ToolHandlerRegistry.createContext(message, "tool", "/workspace", testLogger)
+
+			expect(context.message).toBe(message)
+			expect(context.ask).toBe("tool")
+			expect(context.workspacePath).toBe("/workspace")
+			expect(context.toolInfo).toBeDefined()
+			expect(context.toolInfo?.name).toBe("read_file")
+			expect(context.logger).toBe(testLogger)
+		})
+	})
+
+	describe("custom handlers", () => {
+		it("should accept custom handler list", () => {
+			const customHandler = new DefaultToolHandler()
+			const registry = new ToolHandlerRegistry([customHandler])
+
+			const context = createContext(createToolMessage("any_tool", {}), "command")
+			const handler = registry.getHandler(context)
+
+			expect(handler).toBe(customHandler)
+		})
+	})
+})

apps/cli/src/acp/__tests__/translator.test.ts

@@ -0,0 +1,502 @@
+import type { ClineMessage } from "@roo-code/types"
+
+import {
+	translateToAcpUpdate,
+	parseToolFromMessage,
+	mapToolKind,
+	isPermissionAsk,
+	isCompletionAsk,
+	extractPromptText,
+	extractPromptImages,
+	createPermissionOptions,
+	buildToolCallFromMessage,
+} from "../translator.js"
+
+describe("translateToAcpUpdate", () => {
+	it("should translate text say messages", () => {
+		const message: ClineMessage = {
+			ts: 12345,
+			type: "say",
+			say: "text",
+			text: "Hello, world!",
+		}
+
+		const result = translateToAcpUpdate(message)
+
+		expect(result).toEqual({
+			sessionUpdate: "agent_message_chunk",
+			content: { type: "text", text: "Hello, world!" },
+		})
+	})
+
+	it("should translate reasoning say messages", () => {
+		const message: ClineMessage = {
+			ts: 12345,
+			type: "say",
+			say: "reasoning",
+			text: "I'm thinking about this...",
+		}
+
+		const result = translateToAcpUpdate(message)
+
+		expect(result).toEqual({
+			sessionUpdate: "agent_thought_chunk",
+			content: { type: "text", text: "I'm thinking about this..." },
+		})
+	})
+
+	it("should translate error say messages", () => {
+		const message: ClineMessage = {
+			ts: 12345,
+			type: "say",
+			say: "error",
+			text: "Something went wrong",
+		}
+
+		const result = translateToAcpUpdate(message)
+
+		expect(result).toEqual({
+			sessionUpdate: "agent_message_chunk",
+			content: { type: "text", text: "Error: Something went wrong" },
+		})
+	})
+
+	it("should return null for completion_result", () => {
+		const message: ClineMessage = {
+			ts: 12345,
+			type: "say",
+			say: "completion_result",
+			text: "Task completed",
+		}
+
+		const result = translateToAcpUpdate(message)
+
+		expect(result).toBeNull()
+	})
+
+	it("should return null for ask messages", () => {
+		const message: ClineMessage = {
+			ts: 12345,
+			type: "ask",
+			ask: "tool",
+			text: "Approve this tool?",
+		}
+
+		const result = translateToAcpUpdate(message)
+
+		expect(result).toBeNull()
+	})
+})
+
+describe("parseToolFromMessage", () => {
+	it("should parse JSON tool messages", () => {
+		const message: ClineMessage = {
+			ts: 12345,
+			type: "say",
+			say: "shell_integration_warning",
+			text: JSON.stringify({
+				tool: "read_file",
+				path: "/test/file.txt",
+			}),
+		}
+
+		const result = parseToolFromMessage(message)
+
+		expect(result).not.toBeNull()
+		expect(result?.name).toBe("read_file")
+		// Title is now human-readable based on tool name and filename
+		expect(result?.title).toBe("Read file.txt")
+		expect(result?.locations).toHaveLength(1)
+		expect(result!.locations[0]!.path).toBe("/test/file.txt")
+	})
+
+	it("should extract tool name from text content", () => {
+		const message: ClineMessage = {
+			ts: 12345,
+			type: "say",
+			say: "shell_integration_warning",
+			text: "Using write_file to create the file",
+		}
+
+		const result = parseToolFromMessage(message)
+
+		expect(result).not.toBeNull()
+		expect(result?.name).toBe("write_file")
+	})
+
+	it("should return null for empty text", () => {
+		const message: ClineMessage = {
+			ts: 12345,
+			type: "say",
+			say: "shell_integration_warning",
+			text: "",
+		}
+
+		const result = parseToolFromMessage(message)
+
+		expect(result).toBeNull()
+	})
+})
+
+describe("mapToolKind", () => {
+	it("should map read operations", () => {
+		// Uses exact matching with normalized tool names from TOOL_CATEGORIES
+		expect(mapToolKind("read_file")).toBe("read")
+		expect(mapToolKind("readFile")).toBe("read")
+	})
+
+	it("should map list_files to read kind", () => {
+		// list operations are read-like in the ACP protocol
+		expect(mapToolKind("list_files")).toBe("read")
+		expect(mapToolKind("listFiles")).toBe("read")
+		expect(mapToolKind("listFilesTopLevel")).toBe("read")
+		expect(mapToolKind("listFilesRecursive")).toBe("read")
+	})
+
+	it("should map edit operations", () => {
+		expect(mapToolKind("write_to_file")).toBe("edit")
+		expect(mapToolKind("apply_diff")).toBe("edit")
+		expect(mapToolKind("modify_file")).toBe("edit")
+		expect(mapToolKind("create_file")).toBe("edit")
+		expect(mapToolKind("newFileCreated")).toBe("edit")
+		expect(mapToolKind("editedExistingFile")).toBe("edit")
+	})
+
+	it("should map delete operations", () => {
+		expect(mapToolKind("delete_file")).toBe("delete")
+		expect(mapToolKind("deleteFile")).toBe("delete")
+		expect(mapToolKind("remove_file")).toBe("delete")
+		expect(mapToolKind("removeFile")).toBe("delete")
+	})
+
+	it("should map move operations", () => {
+		expect(mapToolKind("move_file")).toBe("move")
+		expect(mapToolKind("moveFile")).toBe("move")
+		expect(mapToolKind("rename_file")).toBe("move")
+		expect(mapToolKind("renameFile")).toBe("move")
+	})
+
+	it("should map search operations", () => {
+		expect(mapToolKind("search_files")).toBe("search")
+		expect(mapToolKind("searchFiles")).toBe("search")
+		expect(mapToolKind("codebase_search")).toBe("search")
+		expect(mapToolKind("codebaseSearch")).toBe("search")
+		expect(mapToolKind("grep")).toBe("search")
+		expect(mapToolKind("ripgrep")).toBe("search")
+	})
+
+	it("should map execute operations", () => {
+		expect(mapToolKind("execute_command")).toBe("execute")
+		expect(mapToolKind("executeCommand")).toBe("execute")
+		expect(mapToolKind("run_command")).toBe("execute")
+		expect(mapToolKind("runCommand")).toBe("execute")
+	})
+
+	it("should map think operations", () => {
+		expect(mapToolKind("think")).toBe("think")
+		expect(mapToolKind("reason")).toBe("think")
+		expect(mapToolKind("plan")).toBe("think")
+		expect(mapToolKind("analyze")).toBe("think")
+	})
+
+	it("should map fetch operations", () => {
+		// Note: browser_action is NOT mapped to fetch because browser tools are disabled in CLI
+		expect(mapToolKind("fetch")).toBe("fetch")
+		expect(mapToolKind("web_request")).toBe("fetch")
+		expect(mapToolKind("webRequest")).toBe("fetch")
+		expect(mapToolKind("http_get")).toBe("fetch")
+		expect(mapToolKind("httpGet")).toBe("fetch")
+		expect(mapToolKind("http_post")).toBe("fetch")
+		expect(mapToolKind("url_fetch")).toBe("fetch")
+	})
+
+	it("should map browser_action to other (browser tools disabled in CLI)", () => {
+		// browser_action intentionally maps to "other" because browser tools are disabled in CLI mode
+		expect(mapToolKind("browser_action")).toBe("other")
+	})
+
+	it("should map switch_mode operations", () => {
+		expect(mapToolKind("switch_mode")).toBe("switch_mode")
+		expect(mapToolKind("switchMode")).toBe("switch_mode")
+		expect(mapToolKind("set_mode")).toBe("switch_mode")
+		expect(mapToolKind("setMode")).toBe("switch_mode")
+	})
+
+	it("should return other for unknown operations", () => {
+		expect(mapToolKind("unknown_tool")).toBe("other")
+		expect(mapToolKind("custom_operation")).toBe("other")
+		// Tool names that don't exactly match categories also return other
+		expect(mapToolKind("inspect_code")).toBe("other")
+		expect(mapToolKind("get_info")).toBe("other")
+	})
+})
+
+describe("isPermissionAsk", () => {
+	it("should return true for permission-required asks", () => {
+		expect(isPermissionAsk("tool")).toBe(true)
+		expect(isPermissionAsk("command")).toBe(true)
+		expect(isPermissionAsk("browser_action_launch")).toBe(true)
+		expect(isPermissionAsk("use_mcp_server")).toBe(true)
+	})
+
+	it("should return false for other asks", () => {
+		expect(isPermissionAsk("followup")).toBe(false)
+		expect(isPermissionAsk("completion_result")).toBe(false)
+		expect(isPermissionAsk("api_req_failed")).toBe(false)
+	})
+})
+
+describe("isCompletionAsk", () => {
+	it("should return true for completion asks", () => {
+		expect(isCompletionAsk("completion_result")).toBe(true)
+		expect(isCompletionAsk("api_req_failed")).toBe(true)
+		expect(isCompletionAsk("mistake_limit_reached")).toBe(true)
+	})
+
+	it("should return false for other asks", () => {
+		expect(isCompletionAsk("tool")).toBe(false)
+		expect(isCompletionAsk("followup")).toBe(false)
+		expect(isCompletionAsk("command")).toBe(false)
+	})
+})
+
+describe("extractPromptText", () => {
+	it("should extract text from text blocks", () => {
+		const prompt = [
+			{ type: "text" as const, text: "Hello" },
+			{ type: "text" as const, text: "World" },
+		]
+
+		const result = extractPromptText(prompt)
+
+		expect(result).toBe("Hello\nWorld")
+	})
+
+	it("should handle resource_link blocks", () => {
+		const prompt = [
+			{ type: "text" as const, text: "Check this file:" },
+			{
+				type: "resource_link" as const,
+				uri: "file:///test/file.txt",
+				name: "file.txt",
+				mimeType: "text/plain",
+			},
+		]
+
+		const result = extractPromptText(prompt)
+
+		expect(result).toContain("@file:///test/file.txt")
+	})
+
+	it("should handle image blocks", () => {
+		const prompt = [
+			{ type: "text" as const, text: "Look at this:" },
+			{
+				type: "image" as const,
+				data: "base64data",
+				mimeType: "image/png",
+			},
+		]
+
+		const result = extractPromptText(prompt)
+
+		expect(result).toContain("[image content]")
+	})
+})
+
+describe("extractPromptImages", () => {
+	it("should extract image data", () => {
+		const prompt = [
+			{ type: "text" as const, text: "Check this:" },
+			{
+				type: "image" as const,
+				data: "base64data1",
+				mimeType: "image/png",
+			},
+			{
+				type: "image" as const,
+				data: "base64data2",
+				mimeType: "image/jpeg",
+			},
+		]
+
+		const result = extractPromptImages(prompt)
+
+		expect(result).toHaveLength(2)
+		expect(result[0]).toBe("base64data1")
+		expect(result[1]).toBe("base64data2")
+	})
+
+	it("should return empty array when no images", () => {
+		const prompt = [{ type: "text" as const, text: "No images here" }]
+
+		const result = extractPromptImages(prompt)
+
+		expect(result).toHaveLength(0)
+	})
+})
+
+describe("createPermissionOptions", () => {
+	it("should include always allow for tool asks", () => {
+		const options = createPermissionOptions("tool")
+
+		expect(options).toHaveLength(3)
+		expect(options[0]!.optionId).toBe("allow_always")
+		expect(options[0]!.kind).toBe("allow_always")
+	})
+
+	it("should include always allow for command asks", () => {
+		const options = createPermissionOptions("command")
+
+		expect(options).toHaveLength(3)
+		expect(options[0]!.optionId).toBe("allow_always")
+	})
+
+	it("should have basic options for other asks", () => {
+		const options = createPermissionOptions("browser_action_launch")
+
+		expect(options).toHaveLength(2)
+		expect(options[0]!.optionId).toBe("allow")
+		expect(options[1]!.optionId).toBe("reject")
+	})
+})
+
+describe("buildToolCallFromMessage", () => {
+	it("should build a valid tool call", () => {
+		const message: ClineMessage = {
+			ts: 12345,
+			type: "say",
+			say: "shell_integration_warning",
+			text: JSON.stringify({
+				tool: "read_file",
+				path: "/test/file.txt",
+			}),
+		}
+
+		const result = buildToolCallFromMessage(message)
+
+		// Tool ID is deterministic based on message timestamp for debugging
+		expect(result.toolCallId).toBe("tool-12345")
+		// Title is now human-readable based on tool name and filename
+		expect(result.title).toBe("Read file.txt")
+		expect(result.kind).toBe("read")
+		expect(result.status).toBe("pending")
+		expect(result.locations).toHaveLength(1)
+	})
+
+	it("should handle messages without text", () => {
+		const message: ClineMessage = {
+			ts: 12345,
+			type: "say",
+			say: "shell_integration_warning",
+		}
+
+		const result = buildToolCallFromMessage(message)
+
+		// Tool ID is deterministic based on message timestamp for debugging
+		expect(result.toolCallId).toBe("tool-12345")
+		expect(result.kind).toBe("other")
+	})
+
+	it("should not include search path as location for search tools", () => {
+		const message: ClineMessage = {
+			ts: 12345,
+			type: "say",
+			say: "shell_integration_warning",
+			text: JSON.stringify({
+				tool: "searchFiles",
+				path: "src",
+				regex: ".*",
+				filePattern: "*utils*",
+			}),
+		}
+
+		const result = buildToolCallFromMessage(message, "/workspace/project")
+
+		// Search path "src" should NOT become a location
+		expect(result.kind).toBe("search")
+		expect(result.locations).toHaveLength(0)
+	})
+
+	it("should extract file paths from search results content", () => {
+		const message: ClineMessage = {
+			ts: 12345,
+			type: "say",
+			say: "shell_integration_warning",
+			text: JSON.stringify({
+				tool: "search_files",
+				path: "cli",
+				regex: ".*",
+				content:
+					"Found 2 results.\n\n# src/utils/helpers.ts\n  1 | export function helper() {}\n\n# src/components/Button.tsx\n  5 | const Button = () => {}",
+			}),
+		}
+
+		const result = buildToolCallFromMessage(message, "/workspace")
+
+		expect(result.kind).toBe("search")
+		// Should extract file paths from the search results
+		expect(result.locations!).toHaveLength(2)
+		expect(result.locations![0]!.path).toBe("/workspace/src/utils/helpers.ts")
+		expect(result.locations![1]!.path).toBe("/workspace/src/components/Button.tsx")
+	})
+
+	it("should include directory path for list_files tools", () => {
+		const message: ClineMessage = {
+			ts: 12345,
+			type: "say",
+			say: "shell_integration_warning",
+			text: JSON.stringify({
+				tool: "list_files",
+				path: "src/components",
+			}),
+		}
+
+		const result = buildToolCallFromMessage(message, "/workspace")
+
+		expect(result.kind).toBe("read")
+		// Directory path should be included for list_files
+		expect(result.locations!).toHaveLength(1)
+		expect(result.locations![0]!.path).toBe("/workspace/src/components")
+	})
+
+	it("should handle codebase_search tool", () => {
+		const message: ClineMessage = {
+			ts: 12345,
+			type: "say",
+			say: "shell_integration_warning",
+			text: JSON.stringify({
+				tool: "codebase_search",
+				query: "find all utils",
+				path: ".",
+				content: "# lib/utils.js\n  10 | function util() {}",
+			}),
+		}
+
+		const result = buildToolCallFromMessage(message, "/project")
+
+		expect(result.kind).toBe("search")
+		expect(result.locations!).toHaveLength(1)
+		expect(result.locations![0]!.path).toBe("/project/lib/utils.js")
+	})
+
+	it("should deduplicate file paths in search results", () => {
+		const message: ClineMessage = {
+			ts: 12345,
+			type: "say",
+			say: "shell_integration_warning",
+			text: JSON.stringify({
+				tool: "searchFiles",
+				path: "src",
+				content: "# src/file.ts\n  1 | match1\n\n# src/file.ts\n  5 | match2\n\n# src/other.ts\n  3 | match3",
+			}),
+		}
+
+		const result = buildToolCallFromMessage(message, "/workspace")
+
+		// Should deduplicate: src/file.ts appears twice but should only be included once
+		expect(result.locations!).toHaveLength(2)
+		expect(result.locations![0]!.path).toBe("/workspace/src/file.ts")
+		expect(result.locations![1]!.path).toBe("/workspace/src/other.ts")
+	})
+})

apps/cli/src/acp/agent.ts

@@ -0,0 +1,243 @@
+/**
+ * RooCodeAgent
+ *
+ * Implements the ACP Agent interface to expose Roo Code as an ACP-compatible agent.
+ * This allows ACP clients like Zed to use Roo Code as their AI coding assistant.
+ */
+
+import {
+	type Agent,
+	type ClientCapabilities,
+	type CancelNotification,
+	// Requests + Responses
+	type InitializeRequest,
+	type InitializeResponse,
+	type NewSessionRequest,
+	type NewSessionResponse,
+	type SetSessionModeRequest,
+	type SetSessionModeResponse,
+	type SetSessionModelRequest,
+	type SetSessionModelResponse,
+	type AuthenticateRequest,
+	type AuthenticateResponse,
+	type PromptRequest,
+	type PromptResponse,
+	// Classes
+	AgentSideConnection,
+	RequestError,
+	// Constants
+	PROTOCOL_VERSION,
+} from "@agentclientprotocol/sdk"
+import { randomUUID } from "node:crypto"
+
+import { DEFAULT_FLAGS } from "@/types/constants.js"
+import { envVarMap } from "@/lib/utils/provider.js"
+import { login, status } from "@/commands/auth/index.js"
+
+import { AVAILABLE_MODES, DEFAULT_MODELS } from "./types.js"
+import { type AcpSessionOptions, AcpSession } from "./session.js"
+import { acpLog } from "./logger.js"
+import { ModelService, createModelService } from "./model-service.js"
+
+/**
+ * RooCodeAgent implements the ACP Agent interface.
+ *
+ * It manages multiple sessions, each with its own ExtensionHost instance,
+ * and handles protocol-level operations like initialization and authentication.
+ */
+export class RooCodeAgent implements Agent {
+	private sessions: Map<string, AcpSession> = new Map()
+	private clientCapabilities: ClientCapabilities | undefined
+	private isAuthenticated = false
+	private readonly modelService: ModelService
+
+	constructor(
+		private readonly options: AcpSessionOptions,
+		private readonly connection: AgentSideConnection,
+	) {
+		acpLog.info("Agent", `RooCodeAgent constructor: connection=${connection}`)
+		this.modelService = createModelService({ apiKey: options.apiKey })
+	}
+
+	async initialize(params: InitializeRequest): Promise<InitializeResponse> {
+		acpLog.request("initialize", params)
+		this.clientCapabilities = params.clientCapabilities
+
+		// Check if already authenticated via environment or existing credentials.
+		const { authenticated } = await status({ verbose: false })
+		acpLog.debug("Agent", `Auth status: ${authenticated ? "authenticated" : "not authenticated"}`)
+
+		return {
+			protocolVersion: PROTOCOL_VERSION,
+			authMethods: [
+				{
+					id: "roo",
+					name: "Sign in with Roo Code Cloud",
+					description: `Sign in with your Roo Code Cloud account or BYOK by exporting an API key Environment Variable (${Object.values(envVarMap).join(", ")})`,
+				},
+			],
+			agentCapabilities: {
+				loadSession: false,
+				promptCapabilities: {
+					image: true,
+					embeddedContext: true,
+				},
+			},
+		}
+	}
+
+	async newSession(params: NewSessionRequest): Promise<NewSessionResponse> {
+		acpLog.request("newSession", params)
+
+		// @TODO: Detect other env vars for different provider and choose
+		// the correct provider or throw.
+		if (!this.isAuthenticated) {
+			const apiKey = this.options.apiKey || process.env.OPENROUTER_API_KEY
+
+			if (!apiKey) {
+				acpLog.error("Agent", "newSession failed: not authenticated and no API key")
+				throw RequestError.authRequired()
+			}
+
+			this.isAuthenticated = true
+		}
+
+		const sessionId = randomUUID()
+		const provider = this.options.provider || "openrouter"
+		const apiKey = this.options.apiKey || process.env.OPENROUTER_API_KEY
+		const mode = this.options.mode || AVAILABLE_MODES[0]!.id
+		const model = this.options.model || DEFAULT_FLAGS.model
+
+		const session = await AcpSession.create({
+			sessionId,
+			cwd: params.cwd,
+			connection: this.connection,
+			options: {
+				extensionPath: this.options.extensionPath,
+				provider,
+				apiKey,
+				model,
+				mode,
+			},
+			deps: {
+				logger: acpLog,
+			},
+		})
+
+		this.sessions.set(sessionId, session)
+
+		const availableModels = await this.modelService.fetchAvailableModels()
+		const modelExists = availableModels.some((m) => m.modelId === model)
+
+		const response: NewSessionResponse = {
+			sessionId,
+			modes: { currentModeId: mode, availableModes: AVAILABLE_MODES },
+			models: {
+				availableModels,
+				currentModelId: modelExists ? model : DEFAULT_MODELS[0]!.modelId,
+			},
+		}
+
+		acpLog.response("newSession", response)
+		return response
+	}
+
+	async setSessionMode(params: SetSessionModeRequest): Promise<SetSessionModeResponse | void> {
+		acpLog.request("setSessionMode", params)
+		const session = this.sessions.get(params.sessionId)
+
+		if (!session) {
+			acpLog.error("Agent", `setSessionMode failed: session not found: ${params.sessionId}`)
+			throw RequestError.invalidParams(undefined, `Session not found: ${params.sessionId}`)
+		}
+
+		const mode = AVAILABLE_MODES.find((m) => m.id === params.modeId)
+
+		if (!mode) {
+			acpLog.error("Agent", `setSessionMode failed: unknown mode: ${params.modeId}`)
+			throw RequestError.invalidParams(undefined, `Unknown mode: ${params.modeId}`)
+		}
+
+		session.setMode(params.modeId)
+		acpLog.response("setSessionMode", {})
+		return {}
+	}
+
+	async unstable_setSessionModel?(params: SetSessionModelRequest): Promise<SetSessionModelResponse | void> {
+		acpLog.request("setSessionMode", params)
+		const session = this.sessions.get(params.sessionId)
+
+		if (!session) {
+			acpLog.error("Agent", `unstable_setSessionModel failed: session not found: ${params.sessionId}`)
+			throw RequestError.invalidParams(undefined, `Session not found: ${params.sessionId}`)
+		}
+
+		const availableModels = await this.modelService.fetchAvailableModels()
+		const modelExists = availableModels.some((m) => m.modelId === params.modelId)
+
+		if (!modelExists) {
+			acpLog.error("Agent", `unstable_setSessionModel failed: model not found: ${params.modelId}`)
+			throw RequestError.invalidParams(undefined, `Model not found: ${params.modelId}`)
+		}
+
+		session.setModel(params.modelId)
+		acpLog.response("unstable_setSessionModel", {})
+		return {}
+	}
+
+	async authenticate(params: AuthenticateRequest): Promise<AuthenticateResponse | void> {
+		acpLog.request("authenticate", params)
+
+		if (params.methodId !== "roo") {
+			throw RequestError.invalidParams(undefined, `Invalid auth method: ${params.methodId}`)
+		}
+
+		const result = await login({ verbose: false })
+
+		if (!result.success) {
+			throw RequestError.authRequired(undefined, "Failed to authenticate with Roo Code Cloud")
+		}
+
+		this.isAuthenticated = true
+
+		acpLog.response("authenticate", {})
+		return {}
+	}
+
+	async prompt(params: PromptRequest): Promise<PromptResponse> {
+		acpLog.request("prompt", {
+			sessionId: params.sessionId,
+			promptLength: params.prompt?.length ?? 0,
+		})
+
+		const session = this.sessions.get(params.sessionId)
+		if (!session) {
+			acpLog.error("Agent", `prompt failed: session not found: ${params.sessionId}`)
+			throw RequestError.invalidParams(undefined, `Session not found: ${params.sessionId}`)
+		}
+
+		const response = await session.prompt(params)
+		acpLog.response("prompt", response)
+		return response
+	}
+
+	async cancel(params: CancelNotification): Promise<void> {
+		acpLog.request("cancel", { sessionId: params.sessionId })
+
+		const session = this.sessions.get(params.sessionId)
+		if (session) {
+			session.cancel()
+			acpLog.info("Agent", `Cancelled session: ${params.sessionId}`)
+		} else {
+			acpLog.warn("Agent", `cancel: session not found: ${params.sessionId}`)
+		}
+	}
+
+	async dispose(): Promise<void> {
+		acpLog.info("Agent", `Disposing ${this.sessions.size} sessions`)
+		const disposals = Array.from(this.sessions.values()).map((session) => session.dispose())
+		await Promise.all(disposals)
+		this.sessions.clear()
+		acpLog.info("Agent", "All sessions disposed")
+	}
+}

apps/cli/src/acp/command-stream.ts

@@ -0,0 +1,243 @@
+/**
+ * CommandStreamManager
+ *
+ * Manages streaming of command execution output with code fence wrapping.
+ * Handles both live command execution events and final command_output messages.
+ */
+
+import type { ClineMessage } from "@roo-code/types"
+
+import type { IDeltaTracker, IAcpLogger, SendUpdateFn } from "./interfaces.js"
+
+// =============================================================================
+// Types
+// =============================================================================
+
+/**
+ * Information about a pending command execution.
+ */
+export interface PendingCommand {
+	toolCallId: string
+	command: string
+	ts: number
+}
+
+/**
+ * Options for creating a CommandStreamManager.
+ */
+export interface CommandStreamManagerOptions {
+	/** Delta tracker for tracking already-sent content */
+	deltaTracker: IDeltaTracker
+	/** Callback to send session updates */
+	sendUpdate: SendUpdateFn
+	/** Logger instance */
+	logger: IAcpLogger
+}
+
+// =============================================================================
+// CommandStreamManager Class
+// =============================================================================
+
+/**
+ * Manages command output streaming with proper code fence wrapping.
+ *
+ * Responsibilities:
+ * - Track pending command tool calls
+ * - Handle live command execution output (with code fences)
+ * - Handle final command_output messages
+ * - Send tool_call_update notifications
+ */
+export class CommandStreamManager {
+	/**
+	 * Track pending command tool calls for the "Run Command" UI.
+	 * Maps tool call ID to command info.
+	 */
+	private pendingCommandCalls: Map<string, PendingCommand> = new Map()
+
+	/**
+	 * Track which command executions have sent the opening code fence.
+	 * Used to wrap command output in markdown code blocks.
+	 */
+	private commandCodeFencesSent: Set<string> = new Set()
+
+	/**
+	 * Map executionId → toolCallId for robust command output routing.
+	 * The executionId is generated by the extension when the command starts,
+	 * so we establish this mapping when we first see output for an executionId.
+	 * This ensures streaming output goes to the correct tool call, even with
+	 * concurrent commands.
+	 */
+	private executionToToolCallId: Map<string, string> = new Map()
+
+	private readonly deltaTracker: IDeltaTracker
+	private readonly sendUpdate: SendUpdateFn
+	private readonly logger: IAcpLogger
+
+	constructor(options: CommandStreamManagerOptions) {
+		this.deltaTracker = options.deltaTracker
+		this.sendUpdate = options.sendUpdate
+		this.logger = options.logger
+	}
+
+	// ===========================================================================
+	// Public API
+	// ===========================================================================
+
+	/**
+	 * Track a new pending command.
+	 * Called when a command tool call is approved.
+	 */
+	trackCommand(toolCallId: string, command: string, ts: number): void {
+		this.pendingCommandCalls.set(toolCallId, { toolCallId, command, ts })
+	}
+
+	/**
+	 * Handle a command_output message from the extension.
+	 * This handles the final tool_call_update for completion, plus the closing fence.
+	 *
+	 * NOTE: Streaming output is handled by handleExecutionOutput().
+	 * This method handles:
+	 * 1. Sending the closing code fence as agent_message_chunk (if streaming occurred)
+	 * 2. Sending the final tool_call_update with status "completed"
+	 */
+	handleCommandOutput(message: ClineMessage): void {
+		const output = message.text || ""
+		const isPartial = message.partial === true
+
+		// Skip partial updates - streaming is handled by handleExecutionOutput().
+		if (isPartial) {
+			return
+		}
+
+		// Handle completion - update the tool call UI.
+		const pendingCall = this.findMostRecentPendingCommand()
+
+		if (pendingCall) {
+			// Send closing code fence as agent_message_chunk if we had streaming output.
+			const hadStreamingOutput = this.commandCodeFencesSent.has(pendingCall.toolCallId)
+
+			if (hadStreamingOutput) {
+				this.sendUpdate({
+					sessionUpdate: "agent_message_chunk",
+					content: { type: "text", text: "```\n" },
+				})
+
+				this.commandCodeFencesSent.delete(pendingCall.toolCallId)
+			}
+
+			// Command completed - send final tool_call_update with completed status.
+			// Note: Zed doesn't display tool_call_update content, so we just mark it complete.
+			this.sendUpdate({
+				sessionUpdate: "tool_call_update",
+				toolCallId: pendingCall.toolCallId,
+				status: "completed",
+				rawOutput: { output },
+			})
+
+			this.pendingCommandCalls.delete(pendingCall.toolCallId)
+		}
+	}
+
+	/**
+	 * Handle streaming command execution output (live terminal output).
+	 * This provides real-time output during command execution.
+	 *
+	 * Sends output as agent_message_chunk messages for Zed visibility.
+	 * The tool_call UI is updated separately in session-event-handler.
+	 *
+	 * Output is wrapped in markdown code blocks:
+	 * - Opening fence ``` sent on first chunk
+	 * - Subsequent chunks sent as-is (deltas only)
+	 * - Closing fence ``` sent in handleCommandOutput()
+	 *
+	 * Uses executionId → toolCallId mapping for robust routing.
+	 */
+	handleExecutionOutput(executionId: string, output: string): void {
+		// Find or establish the toolCallId for this executionId.
+		let toolCallId = this.executionToToolCallId.get(executionId)
+
+		if (!toolCallId) {
+			// First output for this executionId - establish the mapping.
+			const pendingCall = this.findMostRecentPendingCommand()
+
+			if (!pendingCall) {
+				return
+			}
+
+			toolCallId = pendingCall.toolCallId
+			this.executionToToolCallId.set(executionId, toolCallId)
+		}
+
+		// Use executionId as the message key for delta tracking.
+		const delta = this.deltaTracker.getDelta(executionId, output)
+
+		if (!delta) {
+			return
+		}
+
+		// Send opening code fence on first chunk
+		const isFirstChunk = !this.commandCodeFencesSent.has(toolCallId)
+		if (isFirstChunk) {
+			this.commandCodeFencesSent.add(toolCallId)
+
+			this.sendUpdate({
+				sessionUpdate: "agent_message_chunk",
+				content: { type: "text", text: "```\n" },
+			})
+		}
+
+		// Send the delta as agent_message_chunk for Zed visibility.
+		this.sendUpdate({
+			sessionUpdate: "agent_message_chunk",
+			content: { type: "text", text: delta },
+		})
+	}
+
+	/**
+	 * Check if a message is a command_output message that this manager handles.
+	 */
+	isCommandOutputMessage(message: ClineMessage): boolean {
+		return message.type === "say" && message.say === "command_output"
+	}
+
+	/**
+	 * Reset state for a new prompt.
+	 * Call when starting a new prompt to clear all pending state.
+	 */
+	reset(): void {
+		// Clear all pending commands - any from previous prompts are now stale
+		// and would cause duplicate completion messages if not cleaned up.
+		this.pendingCommandCalls.clear()
+		this.commandCodeFencesSent.clear()
+		this.executionToToolCallId.clear()
+	}
+
+	/**
+	 * Get the number of pending commands (for testing/debugging).
+	 */
+	getPendingCommandCount(): number {
+		return this.pendingCommandCalls.size
+	}
+
+	/**
+	 * Check if there are any open code fences (for testing/debugging).
+	 */
+	hasOpenCodeFences(): boolean {
+		return this.commandCodeFencesSent.size > 0
+	}
+
+	/**
+	 * Find the most recent pending command call.
+	 */
+	private findMostRecentPendingCommand(): PendingCommand | undefined {
+		let pendingCall: PendingCommand | undefined
+
+		for (const [, call] of this.pendingCommandCalls) {
+			if (!pendingCall || call.ts > pendingCall.ts) {
+				pendingCall = call
+			}
+		}
+
+		return pendingCall
+	}
+}

apps/cli/src/acp/content-formatter.ts

@@ -0,0 +1,221 @@
+/**
+ * Content Formatter
+ *
+ * Provides content formatting for ACP UI display.
+ *
+ * This module offers two usage patterns:
+ *
+ * 1. **Direct function imports** (preferred for simple use cases):
+ *    ```ts
+ *    import { formatSearchResults, wrapInCodeBlock } from './content-formatter.js'
+ *    const formatted = wrapInCodeBlock(formatSearchResults(content))
+ *    ```
+ *
+ * 2. **Class-based DI** (for dependency injection in tests):
+ *    ```ts
+ *    import { ContentFormatter, type IContentFormatter } from './content-formatter.js'
+ *    const formatter: IContentFormatter = new ContentFormatter()
+ *    ```
+ */
+
+import type { IContentFormatter } from "./interfaces.js"
+import {
+	formatSearchResults,
+	formatReadContent,
+	wrapInCodeBlock,
+	isUserEcho,
+	readFileContent,
+	readFileContentAsync,
+	extractContentFromParams,
+	type FormatConfig,
+	DEFAULT_FORMAT_CONFIG,
+} from "./utils/index.js"
+import { acpLog } from "./logger.js"
+
+// =============================================================================
+// Direct Exports (Preferred)
+// =============================================================================
+
+// Re-export utility functions for direct use
+export { formatSearchResults, formatReadContent, wrapInCodeBlock, isUserEcho }
+
+// =============================================================================
+// Tool Result Formatting
+// =============================================================================
+
+/**
+ * Format tool result content based on the tool kind.
+ *
+ * Applies appropriate formatting (search summary, truncation, code blocks)
+ * based on the tool type.
+ *
+ * @param kind - The tool kind (search, read, etc.)
+ * @param content - The raw content to format
+ * @param config - Optional formatting configuration
+ * @returns Formatted content
+ */
+export function formatToolResult(kind: string, content: string, config: FormatConfig = DEFAULT_FORMAT_CONFIG): string {
+	switch (kind) {
+		case "search":
+			return wrapInCodeBlock(formatSearchResults(content))
+		case "read":
+			return wrapInCodeBlock(formatReadContent(content, config))
+		default:
+			return content
+	}
+}
+
+/**
+ * Extract file content for readFile operations.
+ *
+ * For readFile tools, the rawInput.content field contains the file PATH
+ * (not the contents), so we need to read the actual file.
+ *
+ * @param rawInput - Tool parameters
+ * @param workspacePath - Workspace path for resolving relative paths
+ * @returns File content or error message, or undefined if no path
+ */
+export function extractFileContent(rawInput: Record<string, unknown>, workspacePath: string): string | undefined {
+	const toolName = (rawInput.tool as string | undefined)?.toLowerCase() || ""
+
+	// Only read file content for readFile tools
+	if (toolName !== "readfile" && toolName !== "read_file") {
+		return extractContentFromParams(rawInput)
+	}
+
+	// Check if we have a path before attempting to read
+	const filePath = rawInput.content as string | undefined
+	const relativePath = rawInput.path as string | undefined
+	if (!filePath && !relativePath) {
+		acpLog.warn("ContentFormatter", "readFile tool has no path")
+		return undefined
+	}
+
+	const result = readFileContent(rawInput, workspacePath)
+	if (result.ok) {
+		acpLog.debug("ContentFormatter", `Read file content: ${result.value.length} chars`)
+		return result.value
+	} else {
+		acpLog.error("ContentFormatter", result.error)
+		return `Error reading file: ${result.error}`
+	}
+}
+
+/**
+ * Extract file content asynchronously for readFile operations.
+ *
+ * @param rawInput - Tool parameters
+ * @param workspacePath - Workspace path for resolving relative paths
+ * @returns Promise with file content or error message
+ */
+export async function extractFileContentAsync(
+	rawInput: Record<string, unknown>,
+	workspacePath: string,
+): Promise<string | undefined> {
+	const toolName = (rawInput.tool as string | undefined)?.toLowerCase() || ""
+
+	// Only read file content for readFile tools
+	if (toolName !== "readfile" && toolName !== "read_file") {
+		return extractContentFromParams(rawInput)
+	}
+
+	// Check if we have a path before attempting to read
+	const filePath = rawInput.content as string | undefined
+	const relativePath = rawInput.path as string | undefined
+	if (!filePath && !relativePath) {
+		acpLog.warn("ContentFormatter", "readFile tool has no path")
+		return undefined
+	}
+
+	const result = await readFileContentAsync(rawInput, workspacePath)
+	if (result.ok) {
+		acpLog.debug("ContentFormatter", `Read file content: ${result.value.length} chars`)
+		return result.value
+	} else {
+		acpLog.error("ContentFormatter", result.error)
+		return `Error reading file: ${result.error}`
+	}
+}
+
+// =============================================================================
+// ContentFormatter Class (for DI)
+// =============================================================================
+
+/**
+ * Formats content for display in the ACP client UI.
+ *
+ * Implements IContentFormatter interface for dependency injection.
+ * For simple use cases, prefer the direct function exports above.
+ *
+ * @example
+ * ```ts
+ * // In production code
+ * const formatter = new ContentFormatter()
+ *
+ * // In tests with mock
+ * const mockFormatter: IContentFormatter = {
+ *   formatToolResult: vi.fn(),
+ *   // ...
+ * }
+ * ```
+ */
+export class ContentFormatter implements IContentFormatter {
+	private readonly config: FormatConfig
+
+	constructor(config?: Partial<FormatConfig>) {
+		this.config = { ...DEFAULT_FORMAT_CONFIG, ...config }
+	}
+
+	formatToolResult(kind: string, content: string): string {
+		return formatToolResult(kind, content, this.config)
+	}
+
+	formatSearchResults(content: string): string {
+		return formatSearchResults(content)
+	}
+
+	formatReadResults(content: string): string {
+		return formatReadContent(content, this.config)
+	}
+
+	wrapInCodeBlock(content: string, language?: string): string {
+		return wrapInCodeBlock(content, language)
+	}
+
+	isUserEcho(text: string, promptText: string | null): boolean {
+		return isUserEcho(text, promptText)
+	}
+
+	/**
+	 * Extract content from rawInput parameters.
+	 * Tries common field names for content.
+	 */
+	extractContentFromRawInput(rawInput: Record<string, unknown>): string | undefined {
+		return extractContentFromParams(rawInput)
+	}
+
+	/**
+	 * Extract file content for readFile operations.
+	 * Delegates to the standalone extractFileContent function.
+	 */
+	extractFileContent(rawInput: Record<string, unknown>, workspacePath: string): string | undefined {
+		return extractFileContent(rawInput, workspacePath)
+	}
+}
+
+// =============================================================================
+// Factory Function
+// =============================================================================
+
+/**
+ * Create a new content formatter with optional configuration.
+ */
+export function createContentFormatter(config?: Partial<FormatConfig>): ContentFormatter {
+	return new ContentFormatter(config)
+}
+
+// =============================================================================
+// Type Exports
+// =============================================================================
+
+export type { FormatConfig as ContentFormatterConfig }

apps/cli/src/acp/delta-tracker.ts

@@ -0,0 +1,71 @@
+/**
+ * DeltaTracker - Utility for computing text deltas
+ *
+ * Tracks what portion of text content has been sent and returns only
+ * the new (delta) portion on subsequent calls. This ensures streaming
+ * content is sent incrementally without duplication.
+ *
+ * @example
+ * ```ts
+ * const tracker = new DeltaTracker()
+ *
+ * tracker.getDelta("msg1", "Hello") // returns "Hello"
+ * tracker.getDelta("msg1", "Hello World") // returns " World"
+ * tracker.getDelta("msg1", "Hello World!") // returns "!"
+ *
+ * tracker.reset() // Clear all tracking for new prompt
+ * ```
+ */
+export class DeltaTracker {
+	private positions: Map<string | number, number> = new Map()
+
+	/**
+	 * Get the delta (new portion) of text that hasn't been sent yet.
+	 * Automatically updates internal tracking when there's new content.
+	 *
+	 * @param id - Unique identifier for the content stream (e.g., message timestamp)
+	 * @param fullText - The full accumulated text so far
+	 * @returns The new portion of text (delta), or empty string if nothing new
+	 */
+	getDelta(id: string | number, fullText: string): string {
+		const lastPos = this.positions.get(id) ?? 0
+		const delta = fullText.slice(lastPos)
+
+		if (delta.length > 0) {
+			this.positions.set(id, fullText.length)
+		}
+
+		return delta
+	}
+
+	/**
+	 * Check if there would be a delta without updating tracking.
+	 * Useful for conditional logic without side effects.
+	 */
+	peekDelta(id: string | number, fullText: string): string {
+		const lastPos = this.positions.get(id) ?? 0
+		return fullText.slice(lastPos)
+	}
+
+	/**
+	 * Reset all tracking. Call when starting a new prompt/session.
+	 */
+	reset(): void {
+		this.positions.clear()
+	}
+
+	/**
+	 * Reset tracking for a specific ID only.
+	 */
+	resetId(id: string | number): void {
+		this.positions.delete(id)
+	}
+
+	/**
+	 * Get the current tracked position for an ID.
+	 * Returns 0 if not tracked.
+	 */
+	getPosition(id: string | number): number {
+		return this.positions.get(id) ?? 0
+	}
+}

apps/cli/src/acp/index.ts

@@ -0,0 +1,2 @@
+export { RooCodeAgent } from "./agent.js"
+export { acpLog } from "./logger.js"

apps/cli/src/acp/interfaces.ts

@@ -0,0 +1,375 @@
+/**
+ * ACP Interfaces
+ *
+ * Defines interfaces for dependency injection and testability.
+ * These interfaces allow for mocking in tests and swapping implementations.
+ */
+
+import type * as acp from "@agentclientprotocol/sdk"
+
+// =============================================================================
+// Logger Interface
+// =============================================================================
+
+/**
+ * Interface for ACP logging.
+ * Allows for different logging implementations (file, console, mock for tests).
+ */
+export interface IAcpLogger {
+	/**
+	 * Log an info message.
+	 */
+	info(component: string, message: string, data?: unknown): void
+
+	/**
+	 * Log a debug message.
+	 */
+	debug(component: string, message: string, data?: unknown): void
+
+	/**
+	 * Log a warning message.
+	 */
+	warn(component: string, message: string, data?: unknown): void
+
+	/**
+	 * Log an error message.
+	 */
+	error(component: string, message: string, data?: unknown): void
+
+	/**
+	 * Log an incoming request.
+	 */
+	request(method: string, params?: unknown): void
+
+	/**
+	 * Log an outgoing response.
+	 */
+	response(method: string, result?: unknown): void
+
+	/**
+	 * Log an outgoing notification.
+	 */
+	notification(method: string, params?: unknown): void
+}
+
+// =============================================================================
+// Content Formatter Interface
+// =============================================================================
+
+/**
+ * Interface for content formatting operations.
+ */
+export interface IContentFormatter {
+	/**
+	 * Format tool result content based on the tool kind.
+	 */
+	formatToolResult(kind: string, content: string): string
+
+	/**
+	 * Format search results into a clean summary with file list.
+	 */
+	formatSearchResults(content: string): string
+
+	/**
+	 * Format read results by truncating long file contents.
+	 */
+	formatReadResults(content: string): string
+
+	/**
+	 * Wrap content in markdown code block for better rendering.
+	 */
+	wrapInCodeBlock(content: string, language?: string): string
+
+	/**
+	 * Check if a text message is an echo of the user's prompt.
+	 */
+	isUserEcho(text: string, promptText: string | null): boolean
+}
+
+// =============================================================================
+// Session Interface
+// =============================================================================
+
+/**
+ * Interface for ACP Session.
+ * Enables mocking for tests.
+ */
+export interface IAcpSession {
+	/**
+	 * Process a prompt request from the ACP client.
+	 */
+	prompt(params: acp.PromptRequest): Promise<acp.PromptResponse>
+
+	/**
+	 * Cancel the current prompt.
+	 */
+	cancel(): void
+
+	/**
+	 * Set the session mode.
+	 */
+	setMode(mode: string): void
+
+	/**
+	 * Dispose of the session and release resources.
+	 */
+	dispose(): Promise<void>
+
+	/**
+	 * Get the session ID.
+	 */
+	getSessionId(): string
+}
+
+// =============================================================================
+// Extension Client Interface
+// =============================================================================
+
+/**
+ * Events emitted by the extension client.
+ */
+export interface ExtensionClientEvents {
+	message: (msg: unknown) => void
+	messageUpdated: (msg: unknown) => void
+	waitingForInput: (event: unknown) => void
+	commandExecutionOutput: (event: unknown) => void
+	taskCompleted: (event: unknown) => void
+}
+
+/**
+ * Interface for extension client interactions.
+ */
+export interface IExtensionClient {
+	on<K extends keyof ExtensionClientEvents>(event: K, handler: ExtensionClientEvents[K]): void
+	off<K extends keyof ExtensionClientEvents>(event: K, handler: ExtensionClientEvents[K]): void
+	respond(text: string): void
+	approve(): void
+	reject(message?: string): void
+}
+
+// =============================================================================
+// Extension Host Interface
+// =============================================================================
+
+/**
+ * Events emitted by the extension host.
+ */
+export interface ExtensionHostEvents {
+	extensionWebviewMessage: (msg: unknown) => void
+}
+
+/**
+ * Interface for extension host interactions.
+ */
+export interface IExtensionHost {
+	/**
+	 * Get the extension client for event handling.
+	 */
+	readonly client: IExtensionClient
+
+	/**
+	 * Subscribe to extension host events.
+	 */
+	on<K extends keyof ExtensionHostEvents>(event: K, handler: ExtensionHostEvents[K]): void
+
+	/**
+	 * Unsubscribe from extension host events.
+	 */
+	off<K extends keyof ExtensionHostEvents>(event: K, handler: ExtensionHostEvents[K]): void
+
+	/**
+	 * Activate the extension host.
+	 */
+	activate(): Promise<void>
+
+	/**
+	 * Dispose of the extension host.
+	 */
+	dispose(): Promise<void>
+
+	/**
+	 * Send a message to the extension.
+	 */
+	sendToExtension(message: unknown): void
+}
+
+// =============================================================================
+// Delta Tracker Interface
+// =============================================================================
+
+/**
+ * Interface for delta tracking.
+ */
+export interface IDeltaTracker {
+	/**
+	 * Get the delta (new portion) of text that hasn't been sent yet.
+	 */
+	getDelta(id: string | number, fullText: string): string
+
+	/**
+	 * Check if there would be a delta without updating tracking.
+	 */
+	peekDelta(id: string | number, fullText: string): string
+
+	/**
+	 * Reset all tracking.
+	 */
+	reset(): void
+
+	/**
+	 * Reset tracking for a specific ID only.
+	 */
+	resetId(id: string | number): void
+}
+
+// =============================================================================
+// Prompt State Interface
+// =============================================================================
+
+/**
+ * Valid states for a prompt turn.
+ *
+ * - idle: No prompt is being processed, ready for new prompts
+ * - processing: A prompt is actively being processed
+ */
+export type PromptStateType = "idle" | "processing"
+
+/**
+ * Result of completing a prompt.
+ */
+export interface PromptCompletionResult {
+	stopReason: acp.StopReason
+}
+
+/**
+ * Interface for prompt state management.
+ */
+export interface IPromptStateMachine {
+	/**
+	 * Get the current state.
+	 */
+	getState(): PromptStateType
+
+	/**
+	 * Get the abort signal for the current prompt.
+	 */
+	getAbortSignal(): AbortSignal | null
+
+	/**
+	 * Get the current prompt text.
+	 */
+	getPromptText(): string | null
+
+	/**
+	 * Check if a prompt can be started.
+	 */
+	canStartPrompt(): boolean
+
+	/**
+	 * Check if currently processing a prompt.
+	 */
+	isProcessing(): boolean
+
+	/**
+	 * Start a new prompt.
+	 */
+	startPrompt(promptText: string): Promise<PromptCompletionResult>
+
+	/**
+	 * Complete the prompt with success or failure.
+	 */
+	complete(success: boolean): acp.StopReason
+
+	/**
+	 * Transition to completion with a specific stop reason.
+	 * This allows direct control over the stop reason (e.g., for cancellation).
+	 */
+	transitionToComplete(stopReason: acp.StopReason): void
+
+	/**
+	 * Cancel the current prompt.
+	 */
+	cancel(): void
+
+	/**
+	 * Reset to idle state.
+	 */
+	reset(): void
+}
+
+// =============================================================================
+// Stream Manager Interfaces
+// =============================================================================
+
+/**
+ * Callback to send an ACP session update.
+ */
+export type SendUpdateFn = (update: acp.SessionNotification["update"]) => void
+
+/**
+ * Options for creating stream managers.
+ */
+export interface StreamManagerOptions {
+	/** Delta tracker for tracking already-sent content */
+	deltaTracker: IDeltaTracker
+	/** Callback to send session updates */
+	sendUpdate: SendUpdateFn
+	/** Logger instance */
+	logger: IAcpLogger
+}
+
+/**
+ * Interface for command output streaming.
+ */
+export interface ICommandStreamManager {
+	trackCommand(toolCallId: string, command: string, ts: number): void
+	handleCommandOutput(message: unknown): void
+	handleExecutionOutput(executionId: string, output: string): void
+	isCommandOutputMessage(message: unknown): boolean
+	reset(): void
+}
+
+/**
+ * Interface for tool content streaming.
+ */
+export interface IToolContentStreamManager {
+	isToolAskMessage(message: unknown): boolean
+	handleToolContentStreaming(message: unknown): boolean
+	reset(): void
+}
+
+// =============================================================================
+// Session Dependencies
+// =============================================================================
+
+/**
+ * Dependencies required for creating an AcpSession.
+ * Enables dependency injection for testing.
+ */
+export interface AcpSessionDependencies {
+	/** Logger instance */
+	logger?: IAcpLogger
+	/** Content formatter instance */
+	contentFormatter?: IContentFormatter
+	/** Delta tracker factory */
+	createDeltaTracker?: () => IDeltaTracker
+	/** Prompt state machine factory */
+	createPromptStateMachine?: () => IPromptStateMachine
+}
+
+// =============================================================================
+// Null/Mock Implementations for Testing
+// =============================================================================
+
+/**
+ * No-op logger implementation for testing.
+ */
+export class NullLogger implements IAcpLogger {
+	info(_component: string, _message: string, _data?: unknown): void {}
+	debug(_component: string, _message: string, _data?: unknown): void {}
+	warn(_component: string, _message: string, _data?: unknown): void {}
+	error(_component: string, _message: string, _data?: unknown): void {}
+	request(_method: string, _params?: unknown): void {}
+	response(_method: string, _result?: unknown): void {}
+	notification(_method: string, _params?: unknown): void {}
+}

apps/cli/src/acp/logger.ts

@@ -0,0 +1,188 @@
+/**
+ * ACP Logger
+ *
+ * Provides file-based logging for ACP debugging.
+ * Logs are written to ~/.roo/acp.log by default.
+ *
+ * Since ACP uses stdin/stdout for protocol communication,
+ * we cannot use console.log for debugging. This logger writes
+ * to a file instead.
+ */
+
+import * as fs from "node:fs"
+import * as path from "node:path"
+import * as os from "node:os"
+
+import type { IAcpLogger } from "./interfaces.js"
+
+// =============================================================================
+// Configuration
+// =============================================================================
+
+const DEFAULT_LOG_DIR = path.join(os.homedir(), ".roo")
+const DEFAULT_LOG_FILE = "acp.log"
+const MAX_LOG_SIZE = 10 * 1024 * 1024 // 10MB
+
+// =============================================================================
+// Logger Class
+// =============================================================================
+
+class AcpLogger implements IAcpLogger {
+	private logPath: string
+	private enabled: boolean = true
+	private stream: fs.WriteStream | null = null
+
+	constructor() {
+		const logDir = process.env.ROO_ACP_LOG_DIR || DEFAULT_LOG_DIR
+		const logFile = process.env.ROO_ACP_LOG_FILE || DEFAULT_LOG_FILE
+		this.logPath = path.join(logDir, logFile)
+
+		// Disable logging if explicitly set to false
+		if (process.env.ROO_ACP_LOG === "false") {
+			this.enabled = false
+		}
+	}
+
+	/**
+	 * Initialize the logger.
+	 * Creates the log directory if it doesn't exist.
+	 */
+	private ensureLogFile(): void {
+		if (!this.enabled) return
+
+		try {
+			const logDir = path.dirname(this.logPath)
+			if (!fs.existsSync(logDir)) {
+				fs.mkdirSync(logDir, { recursive: true })
+			}
+
+			// Rotate log if too large
+			if (fs.existsSync(this.logPath)) {
+				const stats = fs.statSync(this.logPath)
+				if (stats.size > MAX_LOG_SIZE) {
+					const rotatedPath = `${this.logPath}.1`
+					if (fs.existsSync(rotatedPath)) {
+						fs.unlinkSync(rotatedPath)
+					}
+					fs.renameSync(this.logPath, rotatedPath)
+				}
+			}
+
+			// Open stream if not already open
+			if (!this.stream) {
+				this.stream = fs.createWriteStream(this.logPath, { flags: "a" })
+			}
+		} catch (_error) {
+			// Silently disable logging on error
+			this.enabled = false
+		}
+	}
+
+	/**
+	 * Format a log message with timestamp and level.
+	 */
+	private formatMessage(level: string, component: string, message: string, data?: unknown): string {
+		const timestamp = new Date().toISOString()
+		let formatted = `[${timestamp}] [${level}] [${component}] ${message}`
+
+		if (data !== undefined) {
+			try {
+				const dataStr = JSON.stringify(data, null, 2)
+				formatted += `\n${dataStr}`
+			} catch {
+				formatted += ` [Data: unserializable]`
+			}
+		}
+
+		return formatted + "\n"
+	}
+
+	/**
+	 * Write a log entry.
+	 */
+	private write(level: string, component: string, message: string, data?: unknown): void {
+		if (!this.enabled) return
+
+		this.ensureLogFile()
+
+		if (this.stream) {
+			const formatted = this.formatMessage(level, component, message, data)
+			this.stream.write(formatted)
+		}
+	}
+
+	/**
+	 * Log an info message.
+	 */
+	info(component: string, message: string, data?: unknown): void {
+		this.write("INFO", component, message, data)
+	}
+
+	/**
+	 * Log a debug message.
+	 */
+	debug(component: string, message: string, data?: unknown): void {
+		this.write("DEBUG", component, message, data)
+	}
+
+	/**
+	 * Log a warning message.
+	 */
+	warn(component: string, message: string, data?: unknown): void {
+		this.write("WARN", component, message, data)
+	}
+
+	/**
+	 * Log an error message.
+	 */
+	error(component: string, message: string, data?: unknown): void {
+		this.write("ERROR", component, message, data)
+	}
+
+	/**
+	 * Log an incoming request.
+	 */
+	request(method: string, params?: unknown): void {
+		this.write("REQUEST", "ACP", `→ ${method}`, params)
+	}
+
+	/**
+	 * Log an outgoing response.
+	 */
+	response(method: string, result?: unknown): void {
+		this.write("RESPONSE", "ACP", `← ${method}`, result)
+	}
+
+	/**
+	 * Log an outgoing notification.
+	 */
+	notification(method: string, params?: unknown): void {
+		this.write("NOTIFY", "ACP", `→ ${method}`, params)
+	}
+
+	/**
+	 * Get the log file path.
+	 */
+	getLogPath(): string {
+		return this.logPath
+	}
+
+	/**
+	 * Close the logger.
+	 */
+	close(): void {
+		if (this.stream) {
+			this.stream.end()
+			this.stream = null
+		}
+	}
+}
+
+// =============================================================================
+// Singleton Export
+// =============================================================================
+
+export const acpLog = new AcpLogger()
+
+// Log startup
+acpLog.info("Logger", `ACP logging initialized. Log file: ${acpLog.getLogPath()}`)

apps/cli/src/acp/model-service.ts

@@ -0,0 +1,130 @@
+/**
+ * Model Service for ACP
+ *
+ * Fetches and caches available models from the Roo Code API.
+ */
+
+import type { ModelInfo } from "@agentclientprotocol/sdk"
+
+import { DEFAULT_MODELS } from "./types.js"
+import { acpLog } from "./logger.js"
+
+const DEFAULT_API_URL = "https://api.roocode.com"
+const DEFAULT_TIMEOUT = 5_000
+
+interface RooModel {
+	id: string
+	name: string
+	description?: string
+	object?: string
+	created?: number
+	owned_by?: string
+}
+
+export interface ModelServiceOptions {
+	/** Base URL for the API (defaults to DEFAULT_API_URL) */
+	apiUrl?: string
+	/** API key for authentication */
+	apiKey?: string
+	/** Request timeout in milliseconds (defaults to DEFAULT_TIMEOUT) */
+	timeout?: number
+}
+
+/**
+ * Service for fetching and managing available models.
+ */
+export class ModelService {
+	private readonly apiUrl: string
+	private readonly apiKey?: string
+	private readonly timeout: number
+	private cachedModels: ModelInfo[] | null = null
+
+	constructor(options: ModelServiceOptions = {}) {
+		this.apiUrl = options.apiUrl || DEFAULT_API_URL
+		this.apiKey = options.apiKey
+		this.timeout = options.timeout || DEFAULT_TIMEOUT
+	}
+
+	/**
+	 * Fetch available models from the API.
+	 * Returns cached models if available, otherwise fetches from API.
+	 * Falls back to default models on error.
+	 */
+	async fetchAvailableModels(): Promise<ModelInfo[]> {
+		if (this.cachedModels) {
+			return this.cachedModels
+		}
+
+		try {
+			const controller = new AbortController()
+			const timeoutId = setTimeout(() => controller.abort(), this.timeout)
+
+			const headers: Record<string, string> = {
+				"Content-Type": "application/json",
+			}
+
+			if (this.apiKey) {
+				headers["Authorization"] = `Bearer ${this.apiKey}`
+			}
+
+			const response = await fetch(`${this.apiUrl}/proxy/v1/models`, {
+				method: "GET",
+				headers,
+				signal: controller.signal,
+			})
+
+			clearTimeout(timeoutId)
+
+			if (!response.ok) {
+				acpLog.warn("ModelService", `API returned ${response.status}, using default models`)
+				this.cachedModels = DEFAULT_MODELS
+				return this.cachedModels
+			}
+
+			const data = await response.json()
+
+			if (!data.data || !Array.isArray(data.data)) {
+				acpLog.warn("ModelService", "Invalid API response format, using default models")
+				this.cachedModels = DEFAULT_MODELS
+				return this.cachedModels
+			}
+
+			this.cachedModels = this.translateModels(data.data)
+			return this.cachedModels
+		} catch (error) {
+			if (error instanceof Error && error.name === "AbortError") {
+				acpLog.warn("ModelService", "Request timed out, using default models")
+			} else {
+				acpLog.warn(
+					"ModelService",
+					`Failed to fetch models: ${error instanceof Error ? error.message : String(error)}`,
+				)
+			}
+
+			this.cachedModels = DEFAULT_MODELS
+			return this.cachedModels
+		}
+	}
+
+	/**
+	 * Clear the cached models, forcing a refresh on next fetch.
+	 */
+	clearCache(): void {
+		this.cachedModels = null
+	}
+
+	private translateModels(data: RooModel[]): ModelInfo[] {
+		const models: ModelInfo[] = data
+			.map(({ id, name, description }) => ({ modelId: id, name, description }))
+			.sort((a, b) => a.modelId.localeCompare(b.modelId))
+
+		return models.length === 0 ? DEFAULT_MODELS : models
+	}
+}
+
+/**
+ * Create a new ModelService instance.
+ */
+export function createModelService(options?: ModelServiceOptions): ModelService {
+	return new ModelService(options)
+}

apps/cli/src/acp/prompt-state.ts

@@ -0,0 +1,245 @@
+/**
+ * Prompt State Machine
+ *
+ * Manages the lifecycle state of a prompt turn in a type-safe way.
+ * Replaces boolean flags with explicit state transitions and guards.
+ *
+ * State transitions:
+ *   idle -> processing (on startPrompt)
+ *   processing -> idle (on complete/cancel)
+ *   idle -> idle (reset)
+ *
+ * This state machine ensures:
+ * - Only one prompt can be active at a time
+ * - State transitions are valid
+ * - Stop reasons are correctly mapped
+ */
+
+import type * as acp from "@agentclientprotocol/sdk"
+import type { IAcpLogger } from "./interfaces.js"
+import { NullLogger } from "./interfaces.js"
+
+// =============================================================================
+// Types
+// =============================================================================
+
+/**
+ * Valid states for a prompt turn.
+ *
+ * - idle: No prompt is being processed, ready for new prompts
+ * - processing: A prompt is actively being processed
+ */
+export type PromptStateType = "idle" | "processing"
+
+/**
+ * Result of completing a prompt.
+ */
+export interface PromptCompletionResult {
+	stopReason: acp.StopReason
+}
+
+/**
+ * Events that can occur during prompt lifecycle.
+ */
+export type PromptEvent =
+	| { type: "START_PROMPT" }
+	| { type: "COMPLETE"; success: boolean }
+	| { type: "CANCEL" }
+	| { type: "RESET" }
+
+/**
+ * Options for creating a PromptStateMachine.
+ */
+export interface PromptStateMachineOptions {
+	/** Logger instance (optional, defaults to NullLogger) */
+	logger?: IAcpLogger
+}
+
+// =============================================================================
+// PromptStateMachine Class
+// =============================================================================
+
+/**
+ * State machine for managing prompt lifecycle.
+ *
+ * Provides explicit state transitions with validation,
+ * replacing ad-hoc boolean flag management.
+ */
+export class PromptStateMachine {
+	private state: PromptStateType = "idle"
+	private abortController: AbortController | null = null
+	private resolvePrompt: ((result: PromptCompletionResult) => void) | null = null
+	private currentPromptText: string | null = null
+	private readonly logger: IAcpLogger
+
+	constructor(options: PromptStateMachineOptions = {}) {
+		this.logger = options.logger ?? new NullLogger()
+	}
+
+	/**
+	 * Get the current state.
+	 */
+	getState(): PromptStateType {
+		return this.state
+	}
+
+	/**
+	 * Get the abort signal for the current prompt.
+	 */
+	getAbortSignal(): AbortSignal | null {
+		return this.abortController?.signal ?? null
+	}
+
+	/**
+	 * Get the current prompt text (for echo detection).
+	 */
+	getCurrentPromptText(): string | null {
+		return this.currentPromptText
+	}
+
+	/**
+	 * Alias for getCurrentPromptText for compatibility.
+	 */
+	getPromptText(): string | null {
+		return this.currentPromptText
+	}
+
+	/**
+	 * Check if a prompt can be started.
+	 */
+	canStartPrompt(): boolean {
+		return this.state === "idle"
+	}
+
+	/**
+	 * Check if currently processing a prompt.
+	 */
+	isProcessing(): boolean {
+		return this.state === "processing"
+	}
+
+	/**
+	 * Start a new prompt.
+	 *
+	 * @param promptText - The user's prompt text (for echo detection)
+	 * @returns A promise that resolves when the prompt completes
+	 * @throws If a prompt is already in progress
+	 */
+	startPrompt(promptText: string): Promise<PromptCompletionResult> {
+		if (this.state !== "idle") {
+			// Cancel existing prompt first
+			this.cancel()
+		}
+
+		this.state = "processing"
+		this.abortController = new AbortController()
+		this.currentPromptText = promptText
+
+		return new Promise((resolve) => {
+			this.resolvePrompt = resolve
+
+			// Handle abort signal
+			this.abortController?.signal.addEventListener("abort", () => {
+				if (this.state === "processing") {
+					this.transitionToComplete("cancelled")
+				}
+			})
+		})
+	}
+
+	/**
+	 * Complete the prompt with success or failure.
+	 *
+	 * @param success - Whether the task completed successfully
+	 * @returns The stop reason that was used
+	 */
+	complete(success: boolean): acp.StopReason {
+		const stopReason = this.mapSuccessToStopReason(success)
+		this.transitionToComplete(stopReason)
+		return stopReason
+	}
+
+	/**
+	 * Cancel the current prompt.
+	 *
+	 * Safe to call even if no prompt is active.
+	 */
+	cancel(): void {
+		if (this.state !== "processing") {
+			return
+		}
+
+		this.abortController?.abort()
+		// Note: The abort handler will call transitionToComplete
+	}
+
+	/**
+	 * Reset to idle state.
+	 *
+	 * Should be called when starting a new prompt to ensure clean state.
+	 */
+	reset(): void {
+		// Clean up any pending resources
+		if (this.abortController) {
+			this.abortController.abort()
+			this.abortController = null
+		}
+
+		this.state = "idle"
+		this.resolvePrompt = null
+		this.currentPromptText = null
+	}
+
+	// ===========================================================================
+	// Public Methods (for direct control)
+	// ===========================================================================
+
+	/**
+	 * Transition to completion and resolve the promise.
+	 * This is public to allow direct control of the stop reason (e.g., for cancellation).
+	 */
+	transitionToComplete(stopReason: acp.StopReason): void {
+		if (this.state !== "processing") {
+			return
+		}
+
+		this.state = "idle"
+
+		// Resolve the promise
+		if (this.resolvePrompt) {
+			this.resolvePrompt({ stopReason })
+			this.resolvePrompt = null
+		}
+
+		// Clean up
+		this.abortController = null
+		this.currentPromptText = null
+	}
+
+	/**
+	 * Map task success to ACP stop reason.
+	 *
+	 * ACP defines these stop reasons:
+	 * - end_turn: Normal completion
+	 * - max_tokens: Token limit reached
+	 * - max_turn_requests: Request limit reached
+	 * - refusal: Agent refused to continue
+	 * - cancelled: User cancelled
+	 */
+	private mapSuccessToStopReason(success: boolean): acp.StopReason {
+		// Use "refusal" for failed tasks as it's the closest match
+		// (indicates the task couldn't continue normally)
+		return success ? "end_turn" : "refusal"
+	}
+}
+
+// =============================================================================
+// Factory Function
+// =============================================================================
+
+/**
+ * Create a new prompt state machine.
+ */
+export function createPromptStateMachine(options?: PromptStateMachineOptions): PromptStateMachine {
+	return new PromptStateMachine(options)
+}

apps/cli/src/acp/session-event-handler.ts

@@ -0,0 +1,575 @@
+/**
+ * Session Event Handler
+ *
+ * Handles events from the ExtensionClient and ExtensionHost, translating them to ACP updates.
+ */
+
+import type { SessionMode } from "@agentclientprotocol/sdk"
+import type {
+	ClineMessage,
+	ClineAsk,
+	ClineSay,
+	ExtensionMessage,
+	ExtensionState,
+	WebviewMessage,
+	ModeConfig,
+} from "@roo-code/types"
+
+import type { WaitingForInputEvent, TaskCompletedEvent, CommandExecutionOutputEvent } from "@/agent/events.js"
+
+import {
+	translateToAcpUpdate,
+	isPermissionAsk,
+	isCompletionAsk,
+	isTodoListMessage,
+	createPlanUpdateFromMessage,
+} from "./translator.js"
+import { isUserEcho } from "./utils/index.js"
+import type {
+	IAcpLogger,
+	IExtensionClient,
+	IExtensionHost,
+	IPromptStateMachine,
+	ICommandStreamManager,
+	IToolContentStreamManager,
+	IDeltaTracker,
+	SendUpdateFn,
+} from "./interfaces.js"
+import { ToolHandlerRegistry } from "./tool-handler.js"
+
+// =============================================================================
+// Streaming Configuration
+// =============================================================================
+
+/**
+ * Configuration for streaming content types.
+ * Defines which message types should be delta-streamed and how.
+ */
+interface StreamConfig {
+	/** ACP update type to use */
+	readonly updateType: "agent_message_chunk" | "agent_thought_chunk"
+	/** Optional transform to apply to the text before delta tracking */
+	readonly textTransform?: (text: string) => string
+}
+
+/**
+ * Type for the delta stream configuration map.
+ * Uses Partial<Record<ClineSay, StreamConfig>> for type safety.
+ */
+type DeltaStreamConfigMap = Partial<Record<ClineSay, StreamConfig>>
+
+/**
+ * Declarative configuration for which `say` types should be delta-streamed.
+ * Any say type not listed here will fall through to the translator for
+ * non-streaming handling.
+ *
+ * Type safety is enforced by:
+ * - DELTA_STREAM_KEYS constrained to ClineSay values
+ * - DeltaStreamConfigMap type annotation
+ *
+ * To add a new streaming type:
+ * 1. Add the key to DELTA_STREAM_KEYS
+ * 2. Add the configuration below
+ */
+const DELTA_STREAM_CONFIG: DeltaStreamConfigMap = {
+	// Regular text messages from the agent
+	text: { updateType: "agent_message_chunk" },
+
+	// Command output (terminal results, etc.)
+	command_output: { updateType: "agent_message_chunk" },
+
+	// Final completion summary
+	completion_result: { updateType: "agent_message_chunk" },
+
+	// Agent's reasoning/thinking
+	reasoning: { updateType: "agent_thought_chunk" },
+
+	// Error messages (prefixed with "Error: ")
+	error: {
+		updateType: "agent_message_chunk",
+		textTransform: (text: string) => `Error: ${text}`,
+	},
+}
+
+/**
+ * Get stream configuration for a say type.
+ * Returns undefined if the say type is not configured for streaming.
+ */
+function getStreamConfig(sayType: ClineSay): StreamConfig | undefined {
+	return DELTA_STREAM_CONFIG[sayType]
+}
+
+// =============================================================================
+// Types
+// =============================================================================
+
+/**
+ * Dependencies for the SessionEventHandler.
+ */
+export interface SessionEventHandlerDeps {
+	/** Logger instance */
+	logger: IAcpLogger
+	/** Extension client for event subscription */
+	client: IExtensionClient
+	/** Extension host for host-level events (modes, etc.) */
+	extensionHost: IExtensionHost
+	/** Prompt state machine */
+	promptState: IPromptStateMachine
+	/** Delta tracker for streaming */
+	deltaTracker: IDeltaTracker
+	/** Command stream manager */
+	commandStreamManager: ICommandStreamManager
+	/** Tool content stream manager */
+	toolContentStreamManager: IToolContentStreamManager
+	/** Tool handler registry */
+	toolHandlerRegistry: ToolHandlerRegistry
+	/** Callback to send updates */
+	sendUpdate: SendUpdateFn
+	/** Callback to approve extension actions */
+	approveAction: () => void
+	/** Callback to respond with text */
+	respondWithText: (text: string) => void
+	/** Callback to send message to extension */
+	sendToExtension: (message: WebviewMessage) => void
+	/** Workspace path */
+	workspacePath: string
+	/** Initial mode ID */
+	initialModeId: string
+	/** Callback to check if cancellation is in progress */
+	isCancelling: () => boolean
+}
+
+/**
+ * Callback for task completion.
+ */
+export type TaskCompletedCallback = (success: boolean) => void
+
+/**
+ * Callback for mode changes.
+ */
+export type ModeChangedCallback = (modeId: string, availableModes: SessionMode[]) => void
+
+// =============================================================================
+// SessionEventHandler Class
+// =============================================================================
+
+/**
+ * Handles events from the ExtensionClient and ExtensionHost, translating them to ACP updates.
+ *
+ * Responsibilities:
+ * - Subscribe to extension client events
+ * - Subscribe to extension host events (mode changes, etc.)
+ * - Handle streaming for text/reasoning messages
+ * - Handle tool permission requests
+ * - Handle task completion
+ * - Track mode state changes
+ */
+export class SessionEventHandler {
+	private readonly logger: IAcpLogger
+	private readonly client: IExtensionClient
+	private readonly extensionHost: IExtensionHost
+	private readonly promptState: IPromptStateMachine
+	private readonly deltaTracker: IDeltaTracker
+	private readonly commandStreamManager: ICommandStreamManager
+	private readonly toolContentStreamManager: IToolContentStreamManager
+	private readonly toolHandlerRegistry: ToolHandlerRegistry
+	private readonly sendUpdate: SendUpdateFn
+	private readonly approveAction: () => void
+	private readonly respondWithText: (text: string) => void
+	private readonly sendToExtension: (message: WebviewMessage) => void
+	private readonly workspacePath: string
+	private readonly isCancelling: () => boolean
+
+	private taskCompletedCallback: TaskCompletedCallback | null = null
+	private modeChangedCallback: ModeChangedCallback | null = null
+
+	/** Current mode ID (Roo Code mode like 'code', 'architect', etc.) */
+	private currentModeId: string
+
+	/** Available modes from extension state */
+	private availableModes: SessionMode[] = []
+
+	/** Listener for extension host messages */
+	private extensionMessageListener: ((msg: unknown) => void) | null = null
+
+	/**
+	 * Track processed permission requests to prevent duplicates.
+	 * The extension may fire multiple waitingForInput events for the same tool call
+	 * as the message is updated. We deduplicate by generating a stable key from
+	 * the ask type and relevant content.
+	 */
+	private processedPermissions: Set<string> = new Set()
+
+	constructor(deps: SessionEventHandlerDeps) {
+		this.logger = deps.logger
+		this.client = deps.client
+		this.extensionHost = deps.extensionHost
+		this.promptState = deps.promptState
+		this.deltaTracker = deps.deltaTracker
+		this.commandStreamManager = deps.commandStreamManager
+		this.toolContentStreamManager = deps.toolContentStreamManager
+		this.toolHandlerRegistry = deps.toolHandlerRegistry
+		this.sendUpdate = deps.sendUpdate
+		this.approveAction = deps.approveAction
+		this.respondWithText = deps.respondWithText
+		this.sendToExtension = deps.sendToExtension
+		this.workspacePath = deps.workspacePath
+		this.currentModeId = deps.initialModeId
+		this.isCancelling = deps.isCancelling
+	}
+
+	// ===========================================================================
+	// Public API
+	// ===========================================================================
+
+	/**
+	 * Set up event handlers to translate ExtensionClient and ExtensionHost events to ACP updates.
+	 */
+	setupEventHandlers(): void {
+		// Handle new messages
+		this.client.on("message", (msg: unknown) => {
+			this.handleMessage(msg as ClineMessage)
+		})
+
+		// Handle message updates (partial -> complete)
+		this.client.on("messageUpdated", (msg: unknown) => {
+			this.handleMessage(msg as ClineMessage)
+		})
+
+		// Handle permission requests (tool calls, commands, etc.)
+		this.client.on("waitingForInput", (event: unknown) => {
+			void this.handleWaitingForInput(event as WaitingForInputEvent)
+		})
+
+		// Handle streaming command execution output (live terminal output)
+		this.client.on("commandExecutionOutput", (event: unknown) => {
+			const cmdEvent = event as CommandExecutionOutputEvent
+			this.commandStreamManager.handleExecutionOutput(cmdEvent.executionId, cmdEvent.output)
+		})
+
+		// Handle task completion
+		this.client.on("taskCompleted", (event: unknown) => {
+			this.handleTaskCompleted(event as TaskCompletedEvent)
+		})
+
+		// Handle extension host messages (modes, state, etc.)
+		this.extensionMessageListener = (msg: unknown) => {
+			this.handleExtensionMessage(msg as ExtensionMessage)
+		}
+		this.extensionHost.on("extensionWebviewMessage", this.extensionMessageListener)
+	}
+
+	/**
+	 * Set the callback for task completion.
+	 */
+	onTaskCompleted(callback: TaskCompletedCallback): void {
+		this.taskCompletedCallback = callback
+	}
+
+	/**
+	 * Set the callback for mode changes.
+	 */
+	onModeChanged(callback: ModeChangedCallback): void {
+		this.modeChangedCallback = callback
+	}
+
+	/**
+	 * Get the current mode ID.
+	 */
+	getCurrentModeId(): string {
+		return this.currentModeId
+	}
+
+	/**
+	 * Get the available modes.
+	 */
+	getAvailableModes(): SessionMode[] {
+		return this.availableModes
+	}
+
+	/**
+	 * Reset state for a new prompt.
+	 */
+	reset(): void {
+		this.deltaTracker.reset()
+		this.commandStreamManager.reset()
+		this.toolContentStreamManager.reset()
+		this.processedPermissions.clear()
+	}
+
+	/**
+	 * Clean up event listeners.
+	 */
+	cleanup(): void {
+		if (this.extensionMessageListener) {
+			this.extensionHost.off("extensionWebviewMessage", this.extensionMessageListener)
+			this.extensionMessageListener = null
+		}
+	}
+
+	// ===========================================================================
+	// Message Handling
+	// ===========================================================================
+
+	/**
+	 * Handle an incoming message from the extension.
+	 *
+	 * Uses the declarative DELTA_STREAM_CONFIG to automatically determine
+	 * which message types should be delta-streamed and how.
+	 */
+	private handleMessage(message: ClineMessage): void {
+		// Don't process messages if there's no active prompt
+		// NOTE: isCancelling guard REMOVED - we now show all content even during cancellation
+		// so the user can see exactly what was produced before the task paused
+		if (!this.promptState.isProcessing()) {
+			return
+		}
+
+		// === TEST LOGGING: Log messages that arrive during cancellation ===
+		if (this.isCancelling()) {
+			const msgType = message.type === "say" ? `say:${message.say}` : `ask:${message.ask}`
+			const partial = message.partial ? "PARTIAL" : "COMPLETE"
+			this.logger.info("EventHandler", `MSG DURING CANCEL (processing): ${msgType} ${partial} ts=${message.ts}`)
+		}
+
+		// Handle todo list updates - translate to ACP plan updates
+		// Detects both tool asks for updateTodoList and user_edit_todos say messages
+		if (isTodoListMessage(message)) {
+			const planUpdate = createPlanUpdateFromMessage(message)
+			if (planUpdate) {
+				this.sendUpdate(planUpdate)
+			}
+			// Don't return - let the message also be processed by other handlers
+			// (e.g., for permission requests that may follow)
+		}
+
+		// Handle streaming for tool ask messages (file creates/edits)
+		// These contain content that grows as the LLM generates it
+		if (this.toolContentStreamManager.isToolAskMessage(message)) {
+			this.toolContentStreamManager.handleToolContentStreaming(message)
+			return
+		}
+
+		// Check if this is a streaming message type
+		if (message.type === "say" && message.text && message.say) {
+			// Handle command_output specially for the "Run Command" UI
+			if (this.commandStreamManager.isCommandOutputMessage(message)) {
+				this.commandStreamManager.handleCommandOutput(message)
+				return
+			}
+
+			const config = getStreamConfig(message.say)
+
+			if (config) {
+				// Filter out user message echo
+				if (message.say === "text" && isUserEcho(message.text, this.promptState.getPromptText())) {
+					return
+				}
+
+				// Apply text transform if configured (e.g., "Error: " prefix)
+				const textToSend = config.textTransform ? config.textTransform(message.text) : message.text
+
+				// Get delta using the tracker (handles all bookkeeping automatically)
+				const delta = this.deltaTracker.getDelta(message.ts, textToSend)
+
+				if (delta) {
+					this.sendUpdate({
+						sessionUpdate: config.updateType,
+						content: { type: "text", text: delta },
+					})
+				}
+				return
+			}
+		}
+
+		// For non-streaming message types, use the translator
+		const update = translateToAcpUpdate(message)
+		if (update) {
+			this.sendUpdate(update)
+		}
+	}
+
+	// ===========================================================================
+	// Permission Handling
+	// ===========================================================================
+
+	/**
+	 * Handle waiting for input events (permission requests).
+	 */
+	private async handleWaitingForInput(event: WaitingForInputEvent): Promise<void> {
+		const { ask, message } = event
+		const askType = ask as ClineAsk
+
+		// Don't auto-approve asks if there's no active prompt or if cancellation is in progress
+		if (!this.promptState.isProcessing() || this.isCancelling()) {
+			// === TEST LOGGING: Skipped ask due to cancellation ===
+			if (this.isCancelling()) {
+				this.logger.info("EventHandler", `ASK SKIPPED (cancelling): ask=${askType} ts=${message.ts}`)
+			}
+			return
+		}
+
+		// Handle permission-required asks
+		if (isPermissionAsk(askType)) {
+			this.handlePermissionRequest(message, askType)
+			return
+		}
+
+		// Handle completion asks
+		if (isCompletionAsk(askType)) {
+			// Completion is handled by taskCompleted event
+			return
+		}
+
+		// Handle followup questions - auto-continue for now
+		// In a more sophisticated implementation, these could be surfaced
+		// to the ACP client for user input
+		if (askType === "followup") {
+			this.respondWithText("")
+			return
+		}
+
+		// Handle resume_task - auto-resume
+		if (askType === "resume_task") {
+			this.approveAction()
+			return
+		}
+
+		// Handle API failures - auto-retry for now
+		if (askType === "api_req_failed") {
+			this.approveAction()
+			return
+		}
+
+		// Default: approve and continue
+		this.approveAction()
+	}
+
+	/**
+	 * Handle a permission request for a tool call.
+	 *
+	 * Uses the ToolHandlerRegistry for polymorphic dispatch to the appropriate
+	 * handler based on tool type. Auto-approves all tool calls without prompting
+	 * the user, allowing autonomous operation.
+	 *
+	 * For commands, tracks the call to enable the "Run Command" UI with output.
+	 * For other tools (search, read, etc.), both initial and completion updates
+	 * are sent immediately as the results are already available.
+	 */
+	private handlePermissionRequest(message: ClineMessage, ask: ClineAsk): void {
+		// Generate a stable key for deduplication based on ask type and content
+		// The extension may fire multiple waitingForInput events for the same tool
+		// as the message is updated. We use the message text as a stable identifier.
+		const permissionKey = `${ask}:${message.text || ""}`
+
+		// Check if we've already processed this permission request
+		if (this.processedPermissions.has(permissionKey)) {
+			// Still need to approve the action to unblock the extension
+			this.approveAction()
+			return
+		}
+
+		// Mark this permission as processed
+		this.processedPermissions.add(permissionKey)
+
+		// Create context for the tool handler
+		const context = ToolHandlerRegistry.createContext(message, ask, this.workspacePath, this.logger)
+
+		// Dispatch to the appropriate handler via the registry
+		const result = this.toolHandlerRegistry.handle(context)
+
+		// Send the initial in_progress update
+		this.sendUpdate(result.initialUpdate)
+
+		// Track pending commands for the "Run Command" UI
+		if (result.trackAsPendingCommand) {
+			const { toolCallId, command, ts } = result.trackAsPendingCommand
+			this.commandStreamManager.trackCommand(toolCallId, command, ts)
+		}
+
+		// Send completion update if available (non-command tools)
+		if (result.completionUpdate) {
+			this.sendUpdate(result.completionUpdate)
+		}
+
+		// Auto-approve the tool call
+		this.approveAction()
+	}
+
+	// ===========================================================================
+	// Task Completion
+	// ===========================================================================
+
+	/**
+	 * Handle task completion.
+	 */
+	private handleTaskCompleted(event: TaskCompletedEvent): void {
+		if (this.taskCompletedCallback) {
+			this.taskCompletedCallback(event.success)
+		}
+	}
+
+	// ===========================================================================
+	// Extension Message Handling (Modes, State)
+	// ===========================================================================
+
+	/**
+	 * Handle extension messages for mode and state updates.
+	 */
+	private handleExtensionMessage(msg: ExtensionMessage): void {
+		// Handle "modes" message - list of available modes
+		if (msg.type === "modes" && msg.modes) {
+			this.availableModes = msg.modes.map((m) => ({
+				id: m.slug,
+				name: m.name,
+				description: undefined,
+			}))
+		}
+
+		// Handle "state" message - includes current mode
+		if (msg.type === "state" && msg.state) {
+			const state = msg.state as ExtensionState
+			if (state.mode && state.mode !== this.currentModeId) {
+				this.currentModeId = state.mode
+
+				// Send mode update notification
+				this.sendUpdate({
+					sessionUpdate: "current_mode_update",
+					currentModeId: this.currentModeId,
+				})
+
+				// Notify callback
+				if (this.modeChangedCallback) {
+					this.modeChangedCallback(this.currentModeId, this.availableModes)
+				}
+			}
+
+			// Update available modes from customModes
+			if (state.customModes && Array.isArray(state.customModes)) {
+				this.updateAvailableModesFromConfig(state.customModes as ModeConfig[])
+			}
+		}
+	}
+
+	/**
+	 * Update available modes from ModeConfig array.
+	 */
+	private updateAvailableModesFromConfig(modes: ModeConfig[]): void {
+		this.availableModes = modes.map((m) => ({
+			id: m.slug,
+			name: m.name,
+			description: undefined,
+		}))
+	}
+}
+
+// =============================================================================
+// Factory Function
+// =============================================================================
+
+/**
+ * Create a new SessionEventHandler instance.
+ */
+export function createSessionEventHandler(deps: SessionEventHandlerDeps): SessionEventHandler {
+	return new SessionEventHandler(deps)
+}

apps/cli/src/acp/session.ts

@@ -0,0 +1,403 @@
+/**
+ * ACP Session
+ *
+ * Manages a single ACP session, wrapping an ExtensionHost instance.
+ * Handles message translation, event streaming, and permission requests.
+ */
+
+import {
+	type SessionUpdate,
+	type PromptRequest,
+	type PromptResponse,
+	type SessionModeState,
+	AgentSideConnection,
+} from "@agentclientprotocol/sdk"
+
+import type { SupportedProvider } from "@/types/types.js"
+import { getProviderSettings } from "@/lib/utils/provider.js"
+import { type ExtensionHostOptions, ExtensionHost } from "@/agent/extension-host.js"
+import { AgentLoopState } from "@/agent/agent-state.js"
+
+import { DEFAULT_MODELS } from "./types.js"
+import { extractPromptText, extractPromptImages } from "./translator.js"
+import { acpLog } from "./logger.js"
+import { DeltaTracker } from "./delta-tracker.js"
+import { PromptStateMachine } from "./prompt-state.js"
+import { ToolHandlerRegistry } from "./tool-handler.js"
+import { CommandStreamManager } from "./command-stream.js"
+import { ToolContentStreamManager } from "./tool-content-stream.js"
+import { SessionEventHandler, createSessionEventHandler } from "./session-event-handler.js"
+import type {
+	IAcpSession,
+	IAcpLogger,
+	IDeltaTracker,
+	IPromptStateMachine,
+	AcpSessionDependencies,
+} from "./interfaces.js"
+import { type Result, ok, err } from "./utils/index.js"
+
+export interface AcpSessionOptions {
+	extensionPath: string
+	provider: SupportedProvider
+	apiKey?: string
+	model: string
+	mode: string
+}
+
+/**
+ * AcpSession wraps an ExtensionHost instance and bridges it to the ACP protocol.
+ *
+ * Each ACP session creates its own ExtensionHost, which loads the extension
+ * in a sandboxed environment. The session translates events from the
+ * ExtensionClient to ACP session updates and handles permission requests.
+ */
+export class AcpSession implements IAcpSession {
+	/** Logger instance (injected) */
+	private readonly logger: IAcpLogger
+
+	/** State machine for prompt lifecycle management */
+	private readonly promptState: IPromptStateMachine
+
+	/** Delta tracker for streaming content - ensures only new text is sent */
+	private readonly deltaTracker: IDeltaTracker
+
+	/** Tool handler registry for polymorphic tool dispatch */
+	private readonly toolHandlerRegistry: ToolHandlerRegistry
+
+	/** Command stream manager for handling command output */
+	private readonly commandStreamManager: CommandStreamManager
+
+	/** Tool content stream manager for handling file creates/edits */
+	private readonly toolContentStreamManager: ToolContentStreamManager
+
+	/** Session event handler for managing extension events */
+	private readonly eventHandler: SessionEventHandler
+
+	/** Current model ID */
+	private currentModelId: string = DEFAULT_MODELS[0]!.modelId
+
+	/** Track if we're in the process of cancelling a task */
+	private isCancelling: boolean = false
+
+	private constructor(
+		private readonly sessionId: string,
+		private readonly extensionHost: ExtensionHost,
+		private readonly connection: AgentSideConnection,
+		private readonly workspacePath: string,
+		private readonly options: AcpSessionOptions,
+		deps: AcpSessionDependencies = {},
+	) {
+		this.logger = deps.logger ?? acpLog
+		this.promptState = deps.createPromptStateMachine?.() ?? new PromptStateMachine({ logger: this.logger })
+		this.deltaTracker = deps.createDeltaTracker?.() ?? new DeltaTracker()
+
+		const sendUpdate = (update: SessionUpdate) => connection.sessionUpdate({ sessionId: this.sessionId, update })
+
+		this.toolHandlerRegistry = new ToolHandlerRegistry()
+
+		this.commandStreamManager = new CommandStreamManager({
+			deltaTracker: this.deltaTracker,
+			sendUpdate,
+			logger: this.logger,
+		})
+
+		this.toolContentStreamManager = new ToolContentStreamManager({
+			deltaTracker: this.deltaTracker,
+			sendUpdate,
+			logger: this.logger,
+		})
+
+		// Create event handler with extension host for mode tracking.
+		this.eventHandler = createSessionEventHandler({
+			logger: this.logger,
+			client: extensionHost.client,
+			extensionHost,
+			promptState: this.promptState,
+			deltaTracker: this.deltaTracker,
+			commandStreamManager: this.commandStreamManager,
+			toolContentStreamManager: this.toolContentStreamManager,
+			toolHandlerRegistry: this.toolHandlerRegistry,
+			sendUpdate,
+			approveAction: () => this.extensionHost.client.approve(),
+			respondWithText: (text: string, images?: string[]) => this.extensionHost.client.respond(text, images),
+			sendToExtension: (message) => this.extensionHost.sendToExtension(message),
+			workspacePath,
+			initialModeId: this.options.mode,
+			isCancelling: () => this.isCancelling,
+		})
+
+		this.eventHandler.onTaskCompleted((success) => this.handleTaskCompleted(success))
+
+		// Listen for state changes to log and detect cancellation completion.
+		this.extensionHost.client.on("stateChange", (event) => {
+			const prev = event.previousState
+			const curr = event.currentState
+
+			// Only log if something actually changed.
+			const stateChanged =
+				prev.state !== curr.state ||
+				prev.isRunning !== curr.isRunning ||
+				prev.isStreaming !== curr.isStreaming ||
+				prev.currentAsk !== curr.currentAsk
+
+			if (stateChanged) {
+				this.logger.info(
+					"ExtensionClient",
+					`STATE: ${prev.state} → ${curr.state} (running=${curr.isRunning}, streaming=${curr.isStreaming}, ask=${curr.currentAsk || "none"})`,
+				)
+			}
+
+			// If we're cancelling and the extension transitions to NO_TASK or IDLE, complete the cancellation
+			// NO_TASK: messages were cleared
+			// IDLE: task stopped (e.g., completion_result, api_req_failed, or just stopped)
+			if (this.isCancelling) {
+				const newState = curr.state
+				const isTerminalState =
+					newState === AgentLoopState.NO_TASK ||
+					newState === AgentLoopState.IDLE ||
+					newState === AgentLoopState.RESUMABLE
+
+				// Also check if the agent is no longer running/streaming (it has stopped processing)
+				const hasStopped = !curr.isRunning && !curr.isStreaming
+
+				if (isTerminalState || hasStopped) {
+					this.handleCancellationComplete()
+				}
+			}
+		})
+	}
+
+	/**
+	 * Create a new AcpSession.
+	 *
+	 * This initializes an ExtensionHost for the given working directory
+	 * and sets up event handlers to stream updates to the ACP client.
+	 */
+	static async create({
+		sessionId,
+		cwd,
+		connection,
+		options,
+		deps,
+	}: {
+		sessionId: string
+		cwd: string
+		connection: AgentSideConnection
+		options: AcpSessionOptions
+		deps: AcpSessionDependencies
+	}): Promise<AcpSession> {
+		const hostOptions: ExtensionHostOptions = {
+			mode: options.mode,
+			user: null,
+			provider: options.provider as ExtensionHostOptions["provider"],
+			apiKey: options.apiKey,
+			model: options.model,
+			workspacePath: cwd,
+			extensionPath: options.extensionPath,
+			disableOutput: true, // ACP mode: disable direct output, we stream through ACP.
+			ephemeral: true, // Don't persist state - ACP clients manage their own sessions.
+		}
+
+		const extensionHost = new ExtensionHost(hostOptions)
+		await extensionHost.activate()
+
+		const session = new AcpSession(sessionId, extensionHost, connection, cwd, options, deps)
+		session.setupEventHandlers()
+
+		return session
+	}
+
+	// ===========================================================================
+	// Event Handlers
+	// ===========================================================================
+
+	/**
+	 * Set up event handlers to translate ExtensionClient events to ACP updates.
+	 * This includes both ExtensionClient events and ExtensionHost events (modes, state).
+	 */
+	private setupEventHandlers(): void {
+		this.eventHandler.setupEventHandlers()
+	}
+
+	/**
+	 * Reset state for a new prompt.
+	 */
+	private resetForNewPrompt(): void {
+		this.eventHandler.reset()
+		this.isCancelling = false
+	}
+
+	/**
+	 * Handle task completion.
+	 */
+	private handleTaskCompleted(success: boolean): void {
+		// If we're cancelling, override the stop reason to "cancelled"
+		if (this.isCancelling) {
+			this.handleCancellationComplete()
+		} else {
+			// Normal completion
+			this.promptState.complete(success)
+		}
+	}
+
+	/**
+	 * Handle cancellation completion.
+	 * Called when the extension has finished cancelling (either via taskCompleted or NO_TASK transition).
+	 */
+	private handleCancellationComplete(): void {
+		if (!this.isCancelling) {
+			return // Already handled
+		}
+
+		this.isCancelling = false
+
+		// Directly transition to complete with "cancelled" stop reason
+		this.promptState.transitionToComplete("cancelled")
+	}
+
+	// ===========================================================================
+	// ACP Methods
+	// ===========================================================================
+
+	/**
+	 * Process a prompt request from the ACP client.
+	 */
+	async prompt(params: PromptRequest): Promise<PromptResponse> {
+		// Extract text and images from prompt.
+		const text = extractPromptText(params.prompt)
+		const images = extractPromptImages(params.prompt)
+
+		// Check if we're in a resumable state (paused after cancel).
+		// If so, resume the existing conversation instead of starting fresh.
+		const currentState = this.extensionHost.client.getAgentState()
+		if (currentState.state === AgentLoopState.RESUMABLE && currentState.currentAsk === "resume_task") {
+			this.logger.info(
+				"Session",
+				`RESUME TASK: resuming paused task with user input (was ask=${currentState.currentAsk})`,
+			)
+
+			// Reset state for the resumed prompt (but don't cancel - task is already paused)
+			this.eventHandler.reset()
+			this.isCancelling = false
+
+			// Start tracking the prompt
+			const promise = this.promptState.startPrompt(text)
+
+			// Resume the task with the user's message as follow-up
+			this.extensionHost.client.respond(text, images.length > 0 ? images : undefined)
+
+			return promise
+		}
+
+		// Cancel any pending prompt.
+		this.cancel()
+
+		// Reset state for new prompt.
+		this.resetForNewPrompt()
+
+		// Start the prompt using the state machine.
+		const promise = this.promptState.startPrompt(text)
+
+		if (images.length > 0) {
+			this.extensionHost.sendToExtension({ type: "newTask", text, images })
+		} else {
+			this.extensionHost.sendToExtension({ type: "newTask", text })
+		}
+
+		return promise
+	}
+
+	/**
+	 * Cancel the current prompt.
+	 */
+	cancel(): void {
+		if (this.promptState.isProcessing()) {
+			this.isCancelling = true
+			// Content continues flowing to the client during cancellation so users
+			// see what the LLM was generating when cancel was triggered.
+			this.extensionHost.sendToExtension({ type: "cancelTask" })
+			// We wait for the extension to send a taskCompleted event or transition to NO_TASK
+			// which will trigger handleTaskCompleted -> promptState.transitionToComplete("cancelled")
+		}
+	}
+
+	/**
+	 * Set the session mode (Roo Code operational mode like 'code', 'architect').
+	 * The mode change is tracked by the event handler which listens to extension state updates.
+	 */
+	setMode(mode: string): void {
+		this.extensionHost.sendToExtension({ type: "updateSettings", updatedSettings: { mode } })
+	}
+
+	/**
+	 * Set the current model.
+	 * This updates the provider settings to use the specified model.
+	 */
+	setModel(modelId: string): void {
+		this.currentModelId = modelId
+		const updatedSettings = getProviderSettings(this.options.provider, this.options.apiKey, modelId)
+		this.extensionHost.sendToExtension({ type: "updateSettings", updatedSettings })
+	}
+
+	/**
+	 * Get the current mode state (delegated to event handler).
+	 */
+	getModeState(): SessionModeState {
+		return {
+			currentModeId: this.eventHandler.getCurrentModeId(),
+			availableModes: this.eventHandler.getAvailableModes(),
+		}
+	}
+
+	/**
+	 * Get the current mode ID (delegated to event handler).
+	 */
+	getCurrentModeId(): string {
+		return this.eventHandler.getCurrentModeId()
+	}
+
+	/**
+	 * Get the current model ID.
+	 */
+	getCurrentModelId(): string {
+		return this.currentModelId
+	}
+
+	/**
+	 * Dispose of the session and release resources.
+	 */
+	async dispose(): Promise<void> {
+		this.cancel()
+		this.eventHandler.cleanup()
+		await this.extensionHost.dispose()
+	}
+
+	// ===========================================================================
+	// Helpers
+	// ===========================================================================
+
+	/**
+	 * Send an update directly to the ACP client.
+	 *
+	 * @returns Result indicating success or failure with error details.
+	 */
+	private async sendUpdate(update: SessionUpdate): Promise<Result<void>> {
+		try {
+			// Log the update being sent to ACP connection (commented out - too noisy)
+			this.logger.info("Session", `OUT: ${JSON.stringify(update)}`)
+			await this.connection.sessionUpdate({ sessionId: this.sessionId, update })
+			return ok(undefined)
+		} catch (error) {
+			const errorMessage = error instanceof Error ? error.message : String(error)
+			this.logger.error("Session", `Failed to send update: ${errorMessage}`, error)
+			return err(`Failed to send update to ACP client: ${errorMessage}`)
+		}
+	}
+
+	/**
+	 * Get the session ID.
+	 */
+	getSessionId(): string {
+		return this.sessionId
+	}
+}

apps/cli/src/acp/tool-content-stream.ts

@@ -0,0 +1,204 @@
+/**
+ * ToolContentStreamManager
+ *
+ * Manages streaming of tool content (file creates/edits) with headers and code fences.
+ * Provides live feedback as files are being written by the LLM.
+ *
+ * Extracted from session.ts to separate the tool content streaming concern.
+ */
+
+import type { ClineMessage } from "@roo-code/types"
+
+import type { IDeltaTracker, IAcpLogger, SendUpdateFn } from "./interfaces.js"
+import { isFileWriteTool } from "./tool-registry.js"
+import { hasValidFilePath } from "./utils/index.js"
+
+// =============================================================================
+// Types
+// =============================================================================
+
+/**
+ * Options for creating a ToolContentStreamManager.
+ */
+export interface ToolContentStreamManagerOptions {
+	/** Delta tracker for tracking already-sent content */
+	deltaTracker: IDeltaTracker
+	/** Callback to send session updates */
+	sendUpdate: SendUpdateFn
+	/** Logger instance */
+	logger: IAcpLogger
+}
+
+// =============================================================================
+// ToolContentStreamManager Class
+// =============================================================================
+
+/**
+ * Manages streaming of tool content for file creates/edits.
+ *
+ * Responsibilities:
+ * - Track which tools have sent their header
+ * - Stream file content as it's being generated
+ * - Wrap content in proper markdown code blocks
+ * - Clean up tracking state
+ */
+export class ToolContentStreamManager {
+	/**
+	 * Track which tool content streams have sent their header.
+	 * Used to show file path before streaming content.
+	 */
+	private toolContentHeadersSent: Set<number> = new Set()
+
+	private readonly deltaTracker: IDeltaTracker
+	private readonly sendUpdate: SendUpdateFn
+	private readonly logger: IAcpLogger
+
+	constructor(options: ToolContentStreamManagerOptions) {
+		this.deltaTracker = options.deltaTracker
+		this.sendUpdate = options.sendUpdate
+		this.logger = options.logger
+	}
+
+	// ===========================================================================
+	// Public API
+	// ===========================================================================
+
+	/**
+	 * Check if a message is a tool ask message that this manager handles.
+	 */
+	isToolAskMessage(message: ClineMessage): boolean {
+		return message.type === "ask" && message.ask === "tool"
+	}
+
+	/**
+	 * Handle streaming content for tool ask messages (file creates/edits).
+	 *
+	 * This streams the content field from tool JSON as agent_message_chunk updates,
+	 * providing live feedback as files are being written.
+	 *
+	 * @returns true if the message was handled, false if it should fall through
+	 */
+	handleToolContentStreaming(message: ClineMessage): boolean {
+		const isPartial = message.partial === true
+		const ts = message.ts
+		const text = message.text || ""
+
+		// Parse tool info to get the tool name, path, and content
+		const parsed = this.parseToolMessage(text)
+
+		// If we couldn't parse yet (early streaming), skip until we can identify the tool
+		if (!parsed) {
+			return true // Handled (by skipping)
+		}
+
+		const { toolName, toolPath, content } = parsed
+
+		// Only stream content for file write operations (uses tool registry)
+		if (!isFileWriteTool(toolName)) {
+			return true // Handled (by skipping)
+		}
+
+		// Check if we have valid path and content to start streaming
+		// Path must have a file extension to be considered valid (uses shared utility)
+		const validPath = hasValidFilePath(toolPath)
+		const hasContent = content.length > 0
+
+		if (isPartial) {
+			this.handlePartialMessage(ts, toolPath, content, validPath, hasContent)
+		} else {
+			this.handleCompleteMessage(ts, toolPath, content)
+		}
+
+		return true // Handled
+	}
+
+	/**
+	 * Reset state for a new prompt.
+	 */
+	reset(): void {
+		this.toolContentHeadersSent.clear()
+	}
+
+	/**
+	 * Get the number of active headers (for testing/debugging).
+	 */
+	getActiveHeaderCount(): number {
+		return this.toolContentHeadersSent.size
+	}
+
+	// ===========================================================================
+	// Private Methods
+	// ===========================================================================
+
+	/**
+	 * Parse a tool message to extract tool info.
+	 * Returns null if JSON is incomplete (expected early in streaming).
+	 */
+	private parseToolMessage(text: string): { toolName: string; toolPath: string; content: string } | null {
+		try {
+			const toolInfo = JSON.parse(text || "{}") as Record<string, unknown>
+			return {
+				toolName: (toolInfo.tool as string) || "tool",
+				toolPath: (toolInfo.path as string) || "",
+				content: (toolInfo.content as string) || "",
+			}
+		} catch {
+			// Early in streaming, JSON may be incomplete - this is expected
+			return null
+		}
+	}
+
+	/**
+	 * Handle a partial (streaming) tool message.
+	 */
+	private handlePartialMessage(
+		ts: number,
+		toolPath: string,
+		content: string,
+		hasValidPath: boolean,
+		hasContent: boolean,
+	): void {
+		// Send header as soon as we have a valid path (even without content yet)
+		// This provides immediate feedback that a file is being created, reducing
+		// perceived latency during the gap while LLM generates file content.
+		if (hasValidPath && !this.toolContentHeadersSent.has(ts)) {
+			this.toolContentHeadersSent.add(ts)
+			this.sendUpdate({
+				sessionUpdate: "agent_message_chunk",
+				content: { type: "text", text: `\n**Creating ${toolPath}**\n\`\`\`\n` },
+			})
+		}
+
+		// Stream content deltas when content becomes available
+		if (hasValidPath && hasContent) {
+			// Use a unique key for delta tracking: "tool-content-{ts}"
+			const deltaKey = `tool-content-${ts}`
+			const delta = this.deltaTracker.getDelta(deltaKey, content)
+
+			if (delta) {
+				this.sendUpdate({
+					sessionUpdate: "agent_message_chunk",
+					content: { type: "text", text: delta },
+				})
+			}
+		}
+	}
+
+	/**
+	 * Handle a complete (non-partial) tool message.
+	 */
+	private handleCompleteMessage(ts: number, _toolPath: string, _content: string): void {
+		// Message complete - finish streaming and clean up
+		if (this.toolContentHeadersSent.has(ts)) {
+			// Send closing code fence
+			this.sendUpdate({
+				sessionUpdate: "agent_message_chunk",
+				content: { type: "text", text: "\n```\n" },
+			})
+			this.toolContentHeadersSent.delete(ts)
+		}
+
+		// Note: The actual tool_call notification will be sent via handleWaitingForInput
+		// when the waitingForInput event fires (which happens when partial becomes false)
+	}
+}

apps/cli/src/acp/tool-handler.ts

@@ -0,0 +1,480 @@
+/**
+ * Tool Handler Abstraction
+ *
+ * Provides a polymorphic interface for handling different tool types.
+ * Each handler knows how to process a specific category of tool operations,
+ * enabling cleaner separation of concerns and easier testing.
+ */
+
+import type * as acp from "@agentclientprotocol/sdk"
+import type { ClineMessage, ClineAsk } from "@roo-code/types"
+
+import { parseToolFromMessage, type ToolCallInfo } from "./translator.js"
+import type { IAcpLogger } from "./interfaces.js"
+import { isEditTool, isReadTool, isSearchTool, isListFilesTool, mapToolToKind } from "./tool-registry.js"
+import {
+	formatSearchResults,
+	formatReadContent,
+	wrapInCodeBlock,
+	readFileContent,
+	extractContentFromParams,
+	DEFAULT_FORMAT_CONFIG,
+} from "./utils/index.js"
+
+// =============================================================================
+// Types
+// =============================================================================
+
+/**
+ * Context passed to tool handlers for processing.
+ */
+export interface ToolHandlerContext {
+	/** The original message from the extension */
+	message: ClineMessage
+	/** The ask type if this is a permission request */
+	ask: ClineAsk
+	/** Workspace path for resolving relative file paths */
+	workspacePath: string
+	/** Parsed tool information from the message */
+	toolInfo: ToolCallInfo | null
+	/** Logger instance */
+	logger: IAcpLogger
+}
+
+/**
+ * Result of handling a tool call.
+ */
+export interface ToolHandleResult {
+	/** Initial tool_call update to send */
+	initialUpdate: acp.SessionNotification["update"]
+	/** Completion update to send (for non-command tools) */
+	completionUpdate?: acp.SessionNotification["update"]
+	/** Whether to track this as a pending command */
+	trackAsPendingCommand?: {
+		toolCallId: string
+		command: string
+		ts: number
+	}
+}
+
+/**
+ * Interface for tool handlers.
+ *
+ * Each implementation handles a specific category of tools (commands, files, search, etc.)
+ * and knows how to create the appropriate ACP updates.
+ */
+export interface ToolHandler {
+	/**
+	 * Check if this handler can process the given tool.
+	 */
+	canHandle(context: ToolHandlerContext): boolean
+
+	/**
+	 * Handle the tool call and return the appropriate updates.
+	 */
+	handle(context: ToolHandlerContext): ToolHandleResult
+}
+
+// =============================================================================
+// Base Handler
+// =============================================================================
+
+/**
+ * Base class providing common functionality for tool handlers.
+ */
+abstract class BaseToolHandler implements ToolHandler {
+	abstract canHandle(context: ToolHandlerContext): boolean
+	abstract handle(context: ToolHandlerContext): ToolHandleResult
+
+	/**
+	 * Build the basic tool call structure from context.
+	 */
+	protected buildBaseToolCall(context: ToolHandlerContext, kindOverride?: acp.ToolKind): acp.ToolCall {
+		const { message, toolInfo } = context
+
+		return {
+			toolCallId: toolInfo?.id || `tool-${message.ts}`,
+			title: toolInfo?.title || message.text?.slice(0, 100) || "Tool execution",
+			kind: kindOverride ?? (toolInfo ? mapToolToKind(toolInfo.name) : "other"),
+			status: "pending",
+			locations: toolInfo?.locations || [],
+			rawInput: toolInfo?.params || {},
+		}
+	}
+
+	/**
+	 * Create the initial in_progress update.
+	 */
+	protected createInitialUpdate(
+		toolCall: acp.ToolCall,
+		kindOverride?: acp.ToolKind,
+	): acp.SessionNotification["update"] {
+		return {
+			sessionUpdate: "tool_call",
+			...toolCall,
+			kind: kindOverride ?? toolCall.kind,
+			status: "in_progress",
+		}
+	}
+}
+
+// =============================================================================
+// Command Tool Handler
+// =============================================================================
+
+/**
+ * Handles command execution tools.
+ *
+ * Commands are special because:
+ * - They use "execute" kind for the "Run Command" UI
+ * - They track pending calls for output correlation
+ * - Completion comes via command_output messages, not immediately
+ */
+export class CommandToolHandler extends BaseToolHandler {
+	canHandle(context: ToolHandlerContext): boolean {
+		return context.ask === "command"
+	}
+
+	handle(context: ToolHandlerContext): ToolHandleResult {
+		const { message, logger } = context
+
+		const toolCall = this.buildBaseToolCall(context, "execute")
+
+		logger.info("CommandToolHandler", `Handling command: ${toolCall.toolCallId}`)
+
+		return {
+			initialUpdate: this.createInitialUpdate(toolCall, "execute"),
+			trackAsPendingCommand: {
+				toolCallId: toolCall.toolCallId,
+				command: message.text || "",
+				ts: message.ts,
+			},
+		}
+	}
+}
+
+// =============================================================================
+// File Edit Tool Handler
+// =============================================================================
+
+/**
+ * Handles file editing operations (write, apply_diff, create, modify).
+ *
+ * File edits include diff content in the completion update for UI display.
+ */
+export class FileEditToolHandler extends BaseToolHandler {
+	canHandle(context: ToolHandlerContext): boolean {
+		if (context.ask !== "tool") return false
+
+		const toolName = context.toolInfo?.name || ""
+		return isEditTool(toolName)
+	}
+
+	handle(context: ToolHandlerContext): ToolHandleResult {
+		const { toolInfo, logger } = context
+
+		const toolCall = this.buildBaseToolCall(context, "edit")
+
+		// Include diff content if available
+		if (toolInfo?.content && toolInfo.content.length > 0) {
+			toolCall.content = toolInfo.content
+		}
+
+		logger.info("FileEditToolHandler", `Handling file edit: ${toolCall.toolCallId}`)
+
+		const completionUpdate: acp.SessionNotification["update"] = {
+			sessionUpdate: "tool_call_update",
+			toolCallId: toolCall.toolCallId,
+			status: "completed",
+			rawOutput: toolInfo?.params || {},
+		}
+
+		// Include diff content in completion
+		if (toolInfo?.content && toolInfo.content.length > 0) {
+			completionUpdate.content = toolInfo.content
+		}
+
+		return {
+			initialUpdate: this.createInitialUpdate(toolCall, "edit"),
+			completionUpdate,
+		}
+	}
+}
+
+// =============================================================================
+// File Read Tool Handler
+// =============================================================================
+
+/**
+ * Handles file reading operations.
+ *
+ * For readFile tools, the rawInput.content contains the file PATH (not contents),
+ * so we need to read the actual file content.
+ */
+export class FileReadToolHandler extends BaseToolHandler {
+	canHandle(context: ToolHandlerContext): boolean {
+		if (context.ask !== "tool") return false
+
+		const toolName = context.toolInfo?.name || ""
+		return isReadTool(toolName)
+	}
+
+	handle(context: ToolHandlerContext): ToolHandleResult {
+		const { toolInfo, workspacePath, logger } = context
+
+		const toolCall = this.buildBaseToolCall(context, "read")
+		const rawInput = (toolInfo?.params as Record<string, unknown>) || {}
+
+		logger.info("FileReadToolHandler", `Handling file read: ${toolCall.toolCallId}`)
+
+		// Read actual file content using shared utility
+		const result = readFileContent(rawInput, workspacePath)
+		const fileContent = result.ok ? result.value : result.error
+
+		// Format the content (truncate if needed, wrap in code block)
+		const formattedContent = fileContent
+			? wrapInCodeBlock(formatReadContent(fileContent, DEFAULT_FORMAT_CONFIG))
+			: undefined
+
+		const completionUpdate: acp.SessionNotification["update"] = {
+			sessionUpdate: "tool_call_update",
+			toolCallId: toolCall.toolCallId,
+			status: "completed",
+			rawOutput: rawInput,
+		}
+
+		if (formattedContent) {
+			completionUpdate.content = [
+				{
+					type: "content",
+					content: { type: "text", text: formattedContent },
+				},
+			]
+		}
+
+		return {
+			initialUpdate: this.createInitialUpdate(toolCall, "read"),
+			completionUpdate,
+		}
+	}
+}
+
+// =============================================================================
+// Search Tool Handler
+// =============================================================================
+
+/**
+ * Handles search operations (search_files, codebase_search, grep, etc.).
+ *
+ * Search results are formatted into a clean file list with summary.
+ */
+export class SearchToolHandler extends BaseToolHandler {
+	canHandle(context: ToolHandlerContext): boolean {
+		if (context.ask !== "tool") return false
+
+		const toolName = context.toolInfo?.name || ""
+		return isSearchTool(toolName)
+	}
+
+	handle(context: ToolHandlerContext): ToolHandleResult {
+		const { toolInfo, logger } = context
+
+		const toolCall = this.buildBaseToolCall(context, "search")
+		const rawInput = (toolInfo?.params as Record<string, unknown>) || {}
+
+		logger.info("SearchToolHandler", `Handling search: ${toolCall.toolCallId}`)
+
+		// Format search results using shared utility
+		const rawContent = rawInput.content as string | undefined
+		const formattedContent = rawContent ? wrapInCodeBlock(formatSearchResults(rawContent)) : undefined
+
+		const completionUpdate: acp.SessionNotification["update"] = {
+			sessionUpdate: "tool_call_update",
+			toolCallId: toolCall.toolCallId,
+			status: "completed",
+			rawOutput: rawInput,
+		}
+
+		if (formattedContent) {
+			completionUpdate.content = [
+				{
+					type: "content",
+					content: { type: "text", text: formattedContent },
+				},
+			]
+		}
+
+		return {
+			initialUpdate: this.createInitialUpdate(toolCall, "search"),
+			completionUpdate,
+		}
+	}
+}
+
+// =============================================================================
+// List Files Tool Handler
+// =============================================================================
+
+/**
+ * Handles list_files operations.
+ */
+export class ListFilesToolHandler extends BaseToolHandler {
+	canHandle(context: ToolHandlerContext): boolean {
+		if (context.ask !== "tool") return false
+
+		const toolName = context.toolInfo?.name || ""
+		return isListFilesTool(toolName)
+	}
+
+	handle(context: ToolHandlerContext): ToolHandleResult {
+		const { toolInfo, logger } = context
+
+		const toolCall = this.buildBaseToolCall(context, "read")
+		const rawInput = (toolInfo?.params as Record<string, unknown>) || {}
+
+		logger.info("ListFilesToolHandler", `Handling list files: ${toolCall.toolCallId}`)
+
+		// Extract content using shared utility
+		const rawContent = extractContentFromParams(rawInput)
+
+		const completionUpdate: acp.SessionNotification["update"] = {
+			sessionUpdate: "tool_call_update",
+			toolCallId: toolCall.toolCallId,
+			status: "completed",
+			rawOutput: rawInput,
+		}
+
+		if (rawContent) {
+			completionUpdate.content = [
+				{
+					type: "content",
+					content: { type: "text", text: rawContent },
+				},
+			]
+		}
+
+		return {
+			initialUpdate: this.createInitialUpdate(toolCall, "read"),
+			completionUpdate,
+		}
+	}
+}
+
+// =============================================================================
+// Default Tool Handler
+// =============================================================================
+
+/**
+ * Fallback handler for tools not matched by other handlers.
+ */
+export class DefaultToolHandler extends BaseToolHandler {
+	canHandle(_context: ToolHandlerContext): boolean {
+		// Default handler always matches as fallback
+		return true
+	}
+
+	handle(context: ToolHandlerContext): ToolHandleResult {
+		const { toolInfo, logger } = context
+
+		const toolCall = this.buildBaseToolCall(context)
+		const rawInput = (toolInfo?.params as Record<string, unknown>) || {}
+
+		logger.info("DefaultToolHandler", `Handling tool: ${toolCall.toolCallId}, kind: ${toolCall.kind}`)
+
+		// Extract content using shared utility
+		const rawContent = extractContentFromParams(rawInput)
+
+		const completionUpdate: acp.SessionNotification["update"] = {
+			sessionUpdate: "tool_call_update",
+			toolCallId: toolCall.toolCallId,
+			status: "completed",
+			rawOutput: rawInput,
+		}
+
+		if (rawContent) {
+			completionUpdate.content = [
+				{
+					type: "content",
+					content: { type: "text", text: rawContent },
+				},
+			]
+		}
+
+		return {
+			initialUpdate: this.createInitialUpdate(toolCall),
+			completionUpdate,
+		}
+	}
+}
+
+// =============================================================================
+// Tool Handler Registry
+// =============================================================================
+
+/**
+ * Registry that manages tool handlers and dispatches to the appropriate one.
+ *
+ * Handlers are checked in order - the first one that canHandle() returns true wins.
+ * DefaultToolHandler should always be last as it accepts everything.
+ */
+export class ToolHandlerRegistry {
+	private readonly handlers: ToolHandler[]
+
+	constructor(handlers?: ToolHandler[]) {
+		// Default handler order - more specific handlers first
+		this.handlers = handlers || [
+			new CommandToolHandler(),
+			new FileEditToolHandler(),
+			new FileReadToolHandler(),
+			new SearchToolHandler(),
+			new ListFilesToolHandler(),
+			new DefaultToolHandler(),
+		]
+	}
+
+	/**
+	 * Find the appropriate handler for the given context.
+	 */
+	getHandler(context: ToolHandlerContext): ToolHandler {
+		for (const handler of this.handlers) {
+			if (handler.canHandle(context)) {
+				return handler
+			}
+		}
+
+		// Should never happen if DefaultToolHandler is last
+		throw new Error("No handler found for tool - DefaultToolHandler should always match")
+	}
+
+	/**
+	 * Handle a tool call by finding the appropriate handler and dispatching.
+	 */
+	handle(context: ToolHandlerContext): ToolHandleResult {
+		const handler = this.getHandler(context)
+		return handler.handle(context)
+	}
+
+	/**
+	 * Create a context object from message and ask.
+	 */
+	static createContext(
+		message: ClineMessage,
+		ask: ClineAsk,
+		workspacePath: string,
+		logger: IAcpLogger,
+	): ToolHandlerContext {
+		return {
+			message,
+			ask,
+			workspacePath,
+			toolInfo: parseToolFromMessage(message, workspacePath),
+			logger,
+		}
+	}
+}
+
+// =============================================================================
+// Exports
+// =============================================================================
+
+export { BaseToolHandler }

apps/cli/src/acp/tool-registry.ts

@@ -0,0 +1,530 @@
+/**
+ * Tool Registry
+ *
+ * Centralized registry for tool type definitions, categories, and validation schemas.
+ * Provides type-safe tool identification and parameter validation.
+ *
+ * Uses exact matching with normalized tool names to avoid fragile substring matching.
+ */
+
+import { z } from "zod"
+import type * as acp from "@agentclientprotocol/sdk"
+
+// =============================================================================
+// Tool Category Registry Class
+// =============================================================================
+
+/**
+ * Tool category names.
+ */
+export type ToolCategory =
+	| "edit"
+	| "read"
+	| "search"
+	| "list"
+	| "execute"
+	| "delete"
+	| "move"
+	| "think"
+	| "fetch"
+	| "switchMode"
+	| "fileWrite"
+
+/**
+ * Registry for tool categories with automatic Set generation.
+ *
+ * This class ensures that TOOL_CATEGORIES and lookup Sets are always in sync
+ * by generating Sets automatically from the category definitions.
+ */
+class ToolCategoryRegistry {
+	private readonly categories: Map<ToolCategory, Set<string>> = new Map()
+	private readonly toolDefinitions: Record<ToolCategory, readonly string[]>
+
+	constructor() {
+		// Define tool categories with their associated tool names
+		// All tool names are stored in normalized form (lowercase, no separators)
+		this.toolDefinitions = {
+			/** File edit operations (create, write, modify) */
+			edit: [
+				"newfilecreated",
+				"editedexistingfile",
+				"writetofile",
+				"applydiff",
+				"applieddiff",
+				"createfile",
+				"modifyfile",
+			],
+
+			/** File read operations */
+			read: ["readfile"],
+
+			/** File/codebase search operations */
+			search: ["searchfiles", "codebasesearch", "grep", "ripgrep"],
+
+			/** Directory listing operations */
+			list: ["listfiles", "listfilestoplevel", "listfilesrecursive"],
+
+			/** Command/shell execution */
+			execute: ["executecommand", "runcommand"],
+
+			/** File deletion */
+			delete: ["deletefile", "removefile"],
+
+			/** File move/rename */
+			move: ["movefile", "renamefile"],
+
+			/** Reasoning/thinking operations */
+			think: ["think", "reason", "plan", "analyze"],
+
+			/** External fetch/HTTP operations */
+			fetch: ["fetch", "httpget", "httppost", "urlfetch", "webrequest"],
+
+			/** Mode switching operations */
+			switchMode: ["switchmode", "setmode"],
+
+			/** File write operations (for streaming detection) */
+			fileWrite: ["newfilecreated", "writetofile", "createfile", "editedexistingfile", "applydiff", "modifyfile"],
+		}
+
+		// Build Sets automatically from definitions
+		for (const [category, tools] of Object.entries(this.toolDefinitions)) {
+			this.categories.set(category as ToolCategory, new Set(tools))
+		}
+	}
+
+	/**
+	 * Check if a tool name belongs to a specific category.
+	 * Uses O(1) Set lookup.
+	 */
+	isInCategory(toolName: string, category: ToolCategory): boolean {
+		const normalized = this.normalizeToolName(toolName)
+		return this.categories.get(category)?.has(normalized) ?? false
+	}
+
+	/**
+	 * Get all tools in a category.
+	 */
+	getToolsInCategory(category: ToolCategory): readonly string[] {
+		return this.toolDefinitions[category]
+	}
+
+	/**
+	 * Get all category names.
+	 */
+	getCategoryNames(): ToolCategory[] {
+		return Object.keys(this.toolDefinitions) as ToolCategory[]
+	}
+
+	/**
+	 * Normalize a tool name for comparison.
+	 * Converts to lowercase and removes all separators (-, _).
+	 */
+	private normalizeToolName(name: string): string {
+		return name.toLowerCase().replace(/[-_]/g, "")
+	}
+}
+
+// =============================================================================
+// Singleton Registry Instance
+// =============================================================================
+
+/**
+ * Global tool category registry instance.
+ */
+const toolCategoryRegistry = new ToolCategoryRegistry()
+
+// =============================================================================
+// Legacy Exports for Backward Compatibility
+// =============================================================================
+
+/**
+ * Tool categories with their associated tool names.
+ * @deprecated Use toolCategoryRegistry methods instead
+ */
+export const TOOL_CATEGORIES = {
+	edit: toolCategoryRegistry.getToolsInCategory("edit"),
+	read: toolCategoryRegistry.getToolsInCategory("read"),
+	search: toolCategoryRegistry.getToolsInCategory("search"),
+	list: toolCategoryRegistry.getToolsInCategory("list"),
+	execute: toolCategoryRegistry.getToolsInCategory("execute"),
+	delete: toolCategoryRegistry.getToolsInCategory("delete"),
+	move: toolCategoryRegistry.getToolsInCategory("move"),
+	think: toolCategoryRegistry.getToolsInCategory("think"),
+	fetch: toolCategoryRegistry.getToolsInCategory("fetch"),
+	switchMode: toolCategoryRegistry.getToolsInCategory("switchMode"),
+	fileWrite: toolCategoryRegistry.getToolsInCategory("fileWrite"),
+} as const
+
+// =============================================================================
+// Type Definitions
+// =============================================================================
+
+/**
+ * All known tool names (union of all categories)
+ */
+export type KnownToolName = (typeof TOOL_CATEGORIES)[ToolCategory][number]
+
+// =============================================================================
+// Tool Category Detection Functions
+// =============================================================================
+
+/**
+ * Check if a tool name belongs to a specific category using exact matching.
+ * Uses the centralized registry for O(1) lookup.
+ */
+export function isToolInCategory(toolName: string, category: ToolCategory): boolean {
+	return toolCategoryRegistry.isInCategory(toolName, category)
+}
+
+/**
+ * Check if tool is an edit operation.
+ */
+export function isEditTool(toolName: string): boolean {
+	return isToolInCategory(toolName, "edit")
+}
+
+/**
+ * Check if tool is a read operation.
+ */
+export function isReadTool(toolName: string): boolean {
+	return isToolInCategory(toolName, "read")
+}
+
+/**
+ * Check if tool is a search operation.
+ */
+export function isSearchTool(toolName: string): boolean {
+	return isToolInCategory(toolName, "search")
+}
+
+/**
+ * Check if tool is a list files operation.
+ */
+export function isListFilesTool(toolName: string): boolean {
+	return isToolInCategory(toolName, "list")
+}
+
+/**
+ * Check if tool is a command execution operation.
+ */
+export function isExecuteTool(toolName: string): boolean {
+	return isToolInCategory(toolName, "execute")
+}
+
+/**
+ * Check if tool is a delete operation.
+ */
+export function isDeleteTool(toolName: string): boolean {
+	return isToolInCategory(toolName, "delete")
+}
+
+/**
+ * Check if tool is a move/rename operation.
+ */
+export function isMoveTool(toolName: string): boolean {
+	return isToolInCategory(toolName, "move")
+}
+
+/**
+ * Check if tool is a think/reasoning operation.
+ */
+export function isThinkTool(toolName: string): boolean {
+	return isToolInCategory(toolName, "think")
+}
+
+/**
+ * Check if tool is an external fetch operation.
+ */
+export function isFetchTool(toolName: string): boolean {
+	return isToolInCategory(toolName, "fetch")
+}
+
+/**
+ * Check if tool is a mode switching operation.
+ */
+export function isSwitchModeTool(toolName: string): boolean {
+	return isToolInCategory(toolName, "switchMode")
+}
+
+/**
+ * Check if tool is a file write operation (for streaming).
+ */
+export function isFileWriteTool(toolName: string): boolean {
+	return isToolInCategory(toolName, "fileWrite")
+}
+
+// =============================================================================
+// Tool Kind Mapping
+// =============================================================================
+
+/**
+ * Map a tool name to an ACP ToolKind.
+ *
+ * ACP defines these tool kinds for special UI treatment:
+ * - read: Reading files or data
+ * - edit: Modifying files or content
+ * - delete: Removing files or data
+ * - move: Moving or renaming files
+ * - search: Searching for information
+ * - execute: Running commands or code
+ * - think: Internal reasoning or planning
+ * - fetch: Retrieving external data
+ * - switch_mode: Switching the current session mode
+ * - other: Other tool types (default)
+ *
+ * Uses exact category matching for reliability. Falls back to "other" for unknown tools.
+ */
+export function mapToolToKind(toolName: string): acp.ToolKind {
+	// Check exact category matches in priority order
+	// Order matters only for overlapping categories (like fileWrite and edit)
+	if (isToolInCategory(toolName, "switchMode")) {
+		return "switch_mode"
+	}
+	if (isToolInCategory(toolName, "think")) {
+		return "think"
+	}
+	if (isToolInCategory(toolName, "search")) {
+		return "search"
+	}
+	if (isToolInCategory(toolName, "delete")) {
+		return "delete"
+	}
+	if (isToolInCategory(toolName, "move")) {
+		return "move"
+	}
+	if (isToolInCategory(toolName, "edit")) {
+		return "edit"
+	}
+	if (isToolInCategory(toolName, "fetch")) {
+		return "fetch"
+	}
+	if (isToolInCategory(toolName, "read")) {
+		return "read"
+	}
+	if (isToolInCategory(toolName, "list")) {
+		return "read" // list operations are read-like
+	}
+	if (isToolInCategory(toolName, "execute")) {
+		return "execute"
+	}
+
+	// Default to other for unknown tools
+	return "other"
+}
+
+// =============================================================================
+// Zod Schemas for Tool Parameters
+// =============================================================================
+
+/**
+ * Base schema for all tool parameters.
+ */
+const BaseToolParamsSchema = z.object({
+	tool: z.string(),
+})
+
+/**
+ * Schema for file path tools (read, delete, etc.)
+ */
+export const FilePathParamsSchema = BaseToolParamsSchema.extend({
+	path: z.string(),
+	content: z.string().optional(),
+})
+
+/**
+ * Schema for file write/create tools.
+ */
+export const FileWriteParamsSchema = BaseToolParamsSchema.extend({
+	path: z.string(),
+	content: z.string(),
+})
+
+/**
+ * Schema for file move/rename tools.
+ */
+export const FileMoveParamsSchema = BaseToolParamsSchema.extend({
+	path: z.string(),
+	newPath: z.string().optional(),
+	destination: z.string().optional(),
+})
+
+/**
+ * Schema for search tools.
+ */
+export const SearchParamsSchema = BaseToolParamsSchema.extend({
+	path: z.string().optional(),
+	regex: z.string().optional(),
+	query: z.string().optional(),
+	pattern: z.string().optional(),
+	filePattern: z.string().optional(),
+	content: z.string().optional(),
+})
+
+/**
+ * Schema for list files tools.
+ */
+export const ListFilesParamsSchema = BaseToolParamsSchema.extend({
+	path: z.string(),
+	recursive: z.boolean().optional(),
+	content: z.string().optional(),
+})
+
+/**
+ * Schema for command execution tools.
+ */
+export const CommandParamsSchema = BaseToolParamsSchema.extend({
+	command: z.string().optional(),
+	cwd: z.string().optional(),
+})
+
+/**
+ * Schema for think/reasoning tools.
+ */
+export const ThinkParamsSchema = BaseToolParamsSchema.extend({
+	thought: z.string().optional(),
+	reasoning: z.string().optional(),
+	analysis: z.string().optional(),
+})
+
+/**
+ * Schema for mode switching tools.
+ */
+export const SwitchModeParamsSchema = BaseToolParamsSchema.extend({
+	mode: z.string().optional(),
+	modeId: z.string().optional(),
+})
+
+/**
+ * Generic tool params schema (for unknown tools).
+ */
+export const GenericToolParamsSchema = BaseToolParamsSchema.passthrough()
+
+// =============================================================================
+// Parameter Types
+// =============================================================================
+
+export type FilePathParams = z.infer<typeof FilePathParamsSchema>
+export type FileWriteParams = z.infer<typeof FileWriteParamsSchema>
+export type FileMoveParams = z.infer<typeof FileMoveParamsSchema>
+export type SearchParams = z.infer<typeof SearchParamsSchema>
+export type ListFilesParams = z.infer<typeof ListFilesParamsSchema>
+export type CommandParams = z.infer<typeof CommandParamsSchema>
+export type ThinkParams = z.infer<typeof ThinkParamsSchema>
+export type SwitchModeParams = z.infer<typeof SwitchModeParamsSchema>
+export type GenericToolParams = z.infer<typeof GenericToolParamsSchema>
+
+/**
+ * Union of all tool parameter types.
+ */
+export type ToolParams =
+	| FilePathParams
+	| FileWriteParams
+	| FileMoveParams
+	| SearchParams
+	| ListFilesParams
+	| CommandParams
+	| ThinkParams
+	| SwitchModeParams
+	| GenericToolParams
+
+// =============================================================================
+// Parameter Validation
+// =============================================================================
+
+/**
+ * Result of parameter validation.
+ */
+export type ValidationResult<T> = { success: true; data: T } | { success: false; error: z.ZodError }
+
+/**
+ * Validate tool parameters against the appropriate schema.
+ *
+ * @param toolName - Name of the tool
+ * @param params - Raw parameters to validate
+ * @returns Validation result with typed params or error
+ */
+export function validateToolParams(toolName: string, params: unknown): ValidationResult<ToolParams> {
+	// Select schema based on tool category
+	let schema: z.ZodSchema
+
+	if (isEditTool(toolName)) {
+		schema = FileWriteParamsSchema
+	} else if (isReadTool(toolName)) {
+		schema = FilePathParamsSchema
+	} else if (isSearchTool(toolName)) {
+		schema = SearchParamsSchema
+	} else if (isListFilesTool(toolName)) {
+		schema = ListFilesParamsSchema
+	} else if (isExecuteTool(toolName)) {
+		schema = CommandParamsSchema
+	} else if (isDeleteTool(toolName)) {
+		schema = FilePathParamsSchema
+	} else if (isMoveTool(toolName)) {
+		schema = FileMoveParamsSchema
+	} else if (isThinkTool(toolName)) {
+		schema = ThinkParamsSchema
+	} else if (isSwitchModeTool(toolName)) {
+		schema = SwitchModeParamsSchema
+	} else {
+		// Use generic schema for unknown tools
+		schema = GenericToolParamsSchema
+	}
+
+	const result = schema.safeParse(params)
+
+	if (result.success) {
+		return { success: true, data: result.data as ToolParams }
+	}
+
+	return { success: false, error: result.error }
+}
+
+/**
+ * Parse and validate tool parameters, returning undefined on failure.
+ * Use when validation failure should be handled gracefully.
+ *
+ * @param toolName - Name of the tool
+ * @param params - Raw parameters to validate
+ * @returns Validated params or undefined
+ */
+export function parseToolParams(toolName: string, params: unknown): ToolParams | undefined {
+	const result = validateToolParams(toolName, params)
+	return result.success ? result.data : undefined
+}
+
+// =============================================================================
+// Tool Message Parsing
+// =============================================================================
+
+/**
+ * Schema for parsing tool JSON from message text.
+ */
+export const ToolMessageSchema = z
+	.object({
+		tool: z.string(),
+		path: z.string().optional(),
+		content: z.string().optional(),
+	})
+	.passthrough()
+
+export type ToolMessage = z.infer<typeof ToolMessageSchema>
+
+/**
+ * Parse tool information from a JSON message.
+ *
+ * @param text - JSON text to parse
+ * @returns Parsed tool message or undefined if invalid
+ */
+export function parseToolMessage(text: string): ToolMessage | undefined {
+	if (!text.startsWith("{")) {
+		return undefined
+	}
+
+	try {
+		const parsed = JSON.parse(text)
+		const result = ToolMessageSchema.safeParse(parsed)
+		return result.success ? result.data : undefined
+	} catch {
+		return undefined
+	}
+}

apps/cli/src/acp/translator.ts

@@ -0,0 +1,56 @@
+/**
+ * ACP Message Translator
+ *
+ * This file re-exports from the translator/ module for backward compatibility.
+ * The translator has been split into focused modules for better maintainability:
+ *
+ * - translator/diff-parser.ts: Unified diff parsing
+ * - translator/location-extractor.ts: File location extraction
+ * - translator/prompt-extractor.ts: Prompt content extraction
+ * - translator/tool-parser.ts: Tool information parsing
+ * - translator/message-translator.ts: Main message translation
+ * - translator/plan-translator.ts: TodoItem to ACP PlanEntry translation
+ *
+ * Import from this file or directly from translator/index.ts
+ */
+
+// Re-export everything from the translator module
+export {
+	// Diff parsing
+	parseUnifiedDiff,
+	isUnifiedDiff,
+	type ParsedDiff,
+	// Location extraction
+	extractLocations,
+	extractFilePathsFromSearchResults,
+	type LocationParams,
+	// Prompt extraction
+	extractPromptText,
+	extractPromptImages,
+	extractPromptResources,
+	// Tool parsing
+	parseToolFromMessage,
+	generateToolTitle,
+	extractToolContent,
+	buildToolCallFromMessage,
+	type ToolCallInfo,
+	// Message translation
+	translateToAcpUpdate,
+	isPermissionAsk,
+	isCompletionAsk,
+	createPermissionOptions,
+	// Backward compatibility
+	mapToolKind,
+	// Plan translation (TodoItem to ACP PlanEntry)
+	todoItemToPlanEntry,
+	todoListToPlanUpdate,
+	parseTodoListFromMessage,
+	isTodoListMessage,
+	extractTodoListFromMessage,
+	createPlanUpdateFromMessage,
+	type PlanEntry,
+	type PlanEntryPriority,
+	type PlanEntryStatus,
+	type PlanUpdate,
+	type PriorityConfig,
+} from "./translator/index.js"

apps/cli/src/acp/translator/diff-parser.ts

@@ -0,0 +1,106 @@
+/**
+ * Diff Parser
+ *
+ * Parses unified diff format to extract old and new text.
+ * Used for displaying file changes in ACP tool calls.
+ */
+
+// =============================================================================
+// Types
+// =============================================================================
+
+/**
+ * Result of parsing a unified diff.
+ */
+export interface ParsedDiff {
+	/** Original text (null for new files) */
+	oldText: string | null
+	/** New text content */
+	newText: string
+}
+
+// =============================================================================
+// Diff Parsing
+// =============================================================================
+
+/**
+ * Parse a unified diff string to extract old and new text.
+ *
+ * Handles standard unified diff format:
+ * ```
+ * --- a/file.txt
+ * +++ b/file.txt
+ * @@ -1,3 +1,4 @@
+ *  context line
+ * -removed line
+ * +added line
+ *  more context
+ * ```
+ *
+ * For non-diff content (raw file content), returns { oldText: null, newText: content }.
+ *
+ * @param diffString - The diff string to parse
+ * @returns Parsed diff with old and new text, or null if invalid
+ */
+export function parseUnifiedDiff(diffString: string): ParsedDiff | null {
+	if (!diffString) {
+		return null
+	}
+
+	// Check if this is a unified diff format
+	if (!diffString.includes("@@") && !diffString.includes("---") && !diffString.includes("+++")) {
+		// Not a diff, treat as raw content
+		return { oldText: null, newText: diffString }
+	}
+
+	const lines = diffString.split("\n")
+	const oldLines: string[] = []
+	const newLines: string[] = []
+	let inHunk = false
+	let isNewFile = false
+
+	for (const line of lines) {
+		// Check for new file indicator
+		if (line.startsWith("--- /dev/null")) {
+			isNewFile = true
+			continue
+		}
+
+		// Skip diff headers
+		if (line.startsWith("===") || line.startsWith("---") || line.startsWith("+++") || line.startsWith("@@")) {
+			if (line.startsWith("@@")) {
+				inHunk = true
+			}
+			continue
+		}
+
+		if (!inHunk) {
+			continue
+		}
+
+		if (line.startsWith("-")) {
+			// Removed line (old content)
+			oldLines.push(line.slice(1))
+		} else if (line.startsWith("+")) {
+			// Added line (new content)
+			newLines.push(line.slice(1))
+		} else if (line.startsWith(" ") || line === "") {
+			// Context line (in both old and new)
+			const contextLine = line.startsWith(" ") ? line.slice(1) : line
+			oldLines.push(contextLine)
+			newLines.push(contextLine)
+		}
+	}
+
+	return {
+		oldText: isNewFile ? null : oldLines.join("\n") || null,
+		newText: newLines.join("\n"),
+	}
+}
+
+/**
+ * Check if a string appears to be a unified diff.
+ */
+export function isUnifiedDiff(content: string): boolean {
+	return content.includes("@@") || (content.includes("---") && content.includes("+++"))
+}

apps/cli/src/acp/translator/index.ts

@@ -0,0 +1,59 @@
+/**
+ * Translator Module
+ *
+ * Re-exports all translator functionality for backward compatibility.
+ * Import from this module to use the translator features.
+ *
+ * The translator is split into focused modules:
+ * - diff-parser: Unified diff parsing
+ * - location-extractor: File location extraction
+ * - prompt-extractor: Prompt content extraction
+ * - tool-parser: Tool information parsing
+ * - message-translator: Main message translation
+ * - plan-translator: TodoItem to ACP PlanEntry translation
+ */
+
+// Diff parsing
+export { parseUnifiedDiff, isUnifiedDiff, type ParsedDiff } from "./diff-parser.js"
+
+// Location extraction
+export { extractLocations, extractFilePathsFromSearchResults, type LocationParams } from "./location-extractor.js"
+
+// Prompt extraction
+export { extractPromptText, extractPromptImages, extractPromptResources } from "./prompt-extractor.js"
+
+// Tool parsing
+export {
+	parseToolFromMessage,
+	generateToolTitle,
+	extractToolContent,
+	buildToolCallFromMessage,
+	type ToolCallInfo,
+} from "./tool-parser.js"
+
+// Message translation
+export {
+	translateToAcpUpdate,
+	isPermissionAsk,
+	isCompletionAsk,
+	createPermissionOptions,
+} from "./message-translator.js"
+
+// Re-export mapToolKind for backward compatibility
+// (now uses mapToolToKind from tool-registry internally)
+export { mapToolToKind as mapToolKind } from "../tool-registry.js"
+
+// Plan translation (TodoItem to ACP PlanEntry)
+export {
+	todoItemToPlanEntry,
+	todoListToPlanUpdate,
+	parseTodoListFromMessage,
+	isTodoListMessage,
+	extractTodoListFromMessage,
+	createPlanUpdateFromMessage,
+	type PlanEntry,
+	type PlanEntryPriority,
+	type PlanEntryStatus,
+	type PlanUpdate,
+	type PriorityConfig,
+} from "./plan-translator.js"

apps/cli/src/acp/translator/location-extractor.ts

@@ -0,0 +1,136 @@
+/**
+ * Location Extractor
+ *
+ * Extracts file locations from tool parameters for ACP tool calls.
+ * Handles various parameter formats and tool-specific behaviors.
+ */
+
+import type * as acp from "@agentclientprotocol/sdk"
+
+import { isSearchTool, isListFilesTool } from "../tool-registry.js"
+import { resolveFilePathUnsafe } from "../utils/index.js"
+
+// =============================================================================
+// Types
+// =============================================================================
+
+/**
+ * Parameters that may contain file locations.
+ */
+export interface LocationParams {
+	tool?: string
+	path?: string
+	file?: string
+	filePath?: string
+	file_path?: string
+	directory?: string
+	dir?: string
+	paths?: string[]
+	content?: string
+}
+
+// =============================================================================
+// Location Extraction
+// =============================================================================
+
+/**
+ * Extract file locations from tool parameters.
+ *
+ * Handles different tool types:
+ * - Search tools: Extract file paths from search results
+ * - List files: Include the directory being listed
+ * - File operations: Extract path from standard parameters
+ *
+ * @param params - Tool parameters
+ * @param workspacePath - Optional workspace path to resolve relative paths
+ * @returns Array of tool call locations
+ */
+export function extractLocations(params: Record<string, unknown>, workspacePath?: string): acp.ToolCallLocation[] {
+	const locations: acp.ToolCallLocation[] = []
+	const toolName = (params.tool as string | undefined)?.toLowerCase() || ""
+
+	// For search tools, the 'path' parameter is a search scope directory, not a file being accessed.
+	// Don't include it in locations. Instead, try to extract file paths from search results.
+	if (isSearchTool(toolName)) {
+		// Try to extract file paths from search results content
+		const content = params.content as string | undefined
+		if (content) {
+			return extractFilePathsFromSearchResults(content, workspacePath)
+		}
+		return []
+	}
+
+	// For list_files tools, the 'path' is a directory being listed, which is valid to include
+	// but we should mark it as a directory operation rather than a file access
+	if (isListFilesTool(toolName)) {
+		const dirPath = params.path as string | undefined
+		if (dirPath) {
+			const absolutePath = resolveFilePathUnsafe(dirPath, workspacePath)
+			locations.push({ path: absolutePath })
+		}
+		return locations
+	}
+
+	// Check for common path parameters (for file operations)
+	const pathParams = ["path", "file", "filePath", "file_path"]
+	for (const param of pathParams) {
+		if (typeof params[param] === "string") {
+			const filePath = params[param] as string
+			const absolutePath = resolveFilePathUnsafe(filePath, workspacePath)
+			locations.push({ path: absolutePath })
+		}
+	}
+
+	// Check for directory parameters separately (for directory operations)
+	const dirParams = ["directory", "dir"]
+	for (const param of dirParams) {
+		if (typeof params[param] === "string") {
+			const dirPath = params[param] as string
+			const absolutePath = resolveFilePathUnsafe(dirPath, workspacePath)
+			locations.push({ path: absolutePath })
+		}
+	}
+
+	// Check for paths array
+	if (Array.isArray(params.paths)) {
+		for (const p of params.paths) {
+			if (typeof p === "string") {
+				const absolutePath = resolveFilePathUnsafe(p, workspacePath)
+				locations.push({ path: absolutePath })
+			}
+		}
+	}
+
+	return locations
+}
+
+/**
+ * Extract file paths from search results content.
+ *
+ * Search results typically have format: "# path/to/file.ts" for each matched file.
+ *
+ * @param content - Search results content
+ * @param workspacePath - Optional workspace path
+ * @returns Array of locations from search results
+ */
+export function extractFilePathsFromSearchResults(content: string, workspacePath?: string): acp.ToolCallLocation[] {
+	const locations: acp.ToolCallLocation[] = []
+	const seenPaths = new Set<string>()
+
+	// Match file headers in search results (e.g., "# src/utils.ts" or "## path/to/file.js")
+	const fileHeaderPattern = /^#+\s+(.+?\.[a-zA-Z0-9]+)\s*$/gm
+	let match
+
+	while ((match = fileHeaderPattern.exec(content)) !== null) {
+		const filePath = match[1]!.trim()
+		// Skip if we've already seen this path or if it looks like a markdown header (not a file path)
+		if (seenPaths.has(filePath) || (!filePath.includes("/") && !filePath.includes("."))) {
+			continue
+		}
+		seenPaths.add(filePath)
+		const absolutePath = resolveFilePathUnsafe(filePath, workspacePath)
+		locations.push({ path: absolutePath })
+	}
+
+	return locations
+}

apps/cli/src/acp/translator/message-translator.ts

@@ -0,0 +1,179 @@
+/**
+ * Message Translator
+ *
+ * Translates between internal ClineMessage format and ACP protocol format.
+ * This is the main bridge between Roo Code's message system and the ACP protocol.
+ */
+
+import type * as acp from "@agentclientprotocol/sdk"
+import type { ClineMessage, ClineAsk } from "@roo-code/types"
+
+import { mapToolToKind } from "../tool-registry.js"
+import { parseToolFromMessage } from "./tool-parser.js"
+
+// =============================================================================
+// Message to ACP Update Translation
+// =============================================================================
+
+/**
+ * Translate an internal ClineMessage to an ACP session update.
+ * Returns null if the message type should not be sent to ACP.
+ *
+ * @param message - Internal ClineMessage
+ * @returns ACP session update or null
+ */
+export function translateToAcpUpdate(message: ClineMessage): acp.SessionNotification["update"] | null {
+	if (message.type === "say") {
+		switch (message.say) {
+			case "text":
+				// Agent text output
+				return {
+					sessionUpdate: "agent_message_chunk",
+					content: { type: "text", text: message.text || "" },
+				}
+
+			case "reasoning":
+				// Agent reasoning/thinking
+				return {
+					sessionUpdate: "agent_thought_chunk",
+					content: { type: "text", text: message.text || "" },
+				}
+
+			case "shell_integration_warning":
+			case "mcp_server_request_started":
+			case "mcp_server_response":
+				// Tool-related messages
+				return translateToolSayMessage(message)
+
+			case "user_feedback":
+				// User feedback doesn't need to be sent to ACP client
+				return null
+
+			case "error":
+				// Error messages
+				return {
+					sessionUpdate: "agent_message_chunk",
+					content: { type: "text", text: `Error: ${message.text || ""}` },
+				}
+
+			case "completion_result":
+				// Completion is handled at prompt level
+				return null
+
+			case "api_req_started":
+			case "api_req_finished":
+			case "api_req_retried":
+			case "api_req_retry_delayed":
+			case "api_req_deleted":
+				// API request lifecycle events - not sent to ACP
+				return null
+
+			case "command_output":
+				// Command execution - handled through tool_call
+				return null
+
+			default:
+				// Unknown message type
+				return null
+		}
+	}
+
+	// Ask messages are handled separately through permission flow
+	return null
+}
+
+/**
+ * Translate a tool say message to ACP format.
+ *
+ * @param message - Tool-related ClineMessage
+ * @returns ACP session update or null
+ */
+function translateToolSayMessage(message: ClineMessage): acp.SessionNotification["update"] | null {
+	const toolInfo = parseToolFromMessage(message)
+	if (!toolInfo) {
+		return null
+	}
+
+	if (message.partial) {
+		// Tool in progress
+		return {
+			sessionUpdate: "tool_call",
+			toolCallId: toolInfo.id,
+			title: toolInfo.title,
+			kind: mapToolToKind(toolInfo.name),
+			status: "in_progress" as const,
+			locations: toolInfo.locations,
+			rawInput: toolInfo.params,
+		}
+	} else {
+		// Tool completed
+		return {
+			sessionUpdate: "tool_call_update",
+			toolCallId: toolInfo.id,
+			status: "completed" as const,
+			content: [],
+			rawOutput: toolInfo.params,
+		}
+	}
+}
+
+// =============================================================================
+// Ask Type Helpers
+// =============================================================================
+
+/**
+ * Ask types that require permission from the user.
+ */
+const PERMISSION_ASKS: readonly ClineAsk[] = ["tool", "command", "browser_action_launch", "use_mcp_server"]
+
+/**
+ * Check if an ask type requires permission.
+ *
+ * @param ask - The ask type to check
+ * @returns true if permission is required
+ */
+export function isPermissionAsk(ask: ClineAsk): boolean {
+	return PERMISSION_ASKS.includes(ask)
+}
+
+/**
+ * Ask types that indicate task completion.
+ */
+const COMPLETION_ASKS: readonly ClineAsk[] = ["completion_result", "api_req_failed", "mistake_limit_reached"]
+
+/**
+ * Check if an ask type indicates task completion.
+ *
+ * @param ask - The ask type to check
+ * @returns true if this indicates completion
+ */
+export function isCompletionAsk(ask: ClineAsk): boolean {
+	return COMPLETION_ASKS.includes(ask)
+}
+
+// =============================================================================
+// Permission Options
+// =============================================================================
+
+/**
+ * Create standard permission options for a tool call.
+ *
+ * Returns options like "Allow", "Reject", and optionally "Always Allow"
+ * for certain tool types.
+ *
+ * @param ask - The ask type
+ * @returns Array of permission options
+ */
+export function createPermissionOptions(ask: ClineAsk): acp.PermissionOption[] {
+	const baseOptions: acp.PermissionOption[] = [
+		{ optionId: "allow", name: "Allow", kind: "allow_once" },
+		{ optionId: "reject", name: "Reject", kind: "reject_once" },
+	]
+
+	// Add "allow always" option for certain ask types
+	if (ask === "tool" || ask === "command") {
+		return [{ optionId: "allow_always", name: "Always Allow", kind: "allow_always" }, ...baseOptions]
+	}
+
+	return baseOptions
+}

apps/cli/src/acp/translator/plan-translator.ts

@@ -0,0 +1,260 @@
+/**
+ * Plan Translator
+ *
+ * Translates between Roo CLI TodoItem format and ACP PlanEntry format.
+ * This enables the agent to communicate execution plans to ACP clients
+ * when using the update_todo_list tool.
+ *
+ * @see https://agentclientprotocol.com/protocol/agent-plan
+ */
+
+import type { TodoItem } from "@roo-code/types"
+
+// =============================================================================
+// Types
+// =============================================================================
+
+/**
+ * Priority levels for plan entries.
+ * Maps to ACP PlanEntryPriority.
+ */
+export type PlanEntryPriority = "high" | "medium" | "low"
+
+/**
+ * Status levels for plan entries.
+ * Maps to ACP PlanEntryStatus (same as TodoStatus).
+ */
+export type PlanEntryStatus = "pending" | "in_progress" | "completed"
+
+/**
+ * A single entry in the execution plan.
+ * Represents a task or goal that the agent intends to accomplish.
+ */
+export interface PlanEntry {
+	/** Human-readable description of what this task aims to accomplish */
+	content: string
+	/** The relative importance of this task */
+	priority: PlanEntryPriority
+	/** Current execution status of this task */
+	status: PlanEntryStatus
+}
+
+/**
+ * ACP Plan session update payload.
+ */
+export interface PlanUpdate {
+	sessionUpdate: "plan"
+	entries: PlanEntry[]
+}
+
+/**
+ * Configuration for priority assignment when converting todos to plan entries.
+ */
+export interface PriorityConfig {
+	/** Default priority for all items (default: "medium") */
+	defaultPriority: PlanEntryPriority
+	/** Assign high priority to in_progress items (default: true) */
+	prioritizeInProgress: boolean
+	/** Assign higher priority to earlier items in the list (default: false) */
+	prioritizeByOrder: boolean
+	/** Number of top items to mark as high priority when prioritizeByOrder is true */
+	highPriorityCount: number
+}
+
+/**
+ * Default priority configuration.
+ */
+const DEFAULT_PRIORITY_CONFIG: PriorityConfig = {
+	defaultPriority: "medium",
+	prioritizeInProgress: true,
+	prioritizeByOrder: false,
+	highPriorityCount: 3,
+}
+
+// =============================================================================
+// Priority Determination
+// =============================================================================
+
+/**
+ * Determine the priority of a todo item based on configuration.
+ *
+ * @param item - The todo item
+ * @param index - Position in the list (0-based)
+ * @param total - Total number of items
+ * @param config - Priority configuration
+ * @returns The determined priority
+ */
+function determinePriority(item: TodoItem, index: number, total: number, config: PriorityConfig): PlanEntryPriority {
+	// In-progress items get high priority
+	if (config.prioritizeInProgress && item.status === "in_progress") {
+		return "high"
+	}
+
+	// Order-based priority
+	if (config.prioritizeByOrder && total > 0) {
+		if (index < config.highPriorityCount) {
+			return "high"
+		}
+		if (index < Math.floor(total / 2)) {
+			return "medium"
+		}
+		return "low"
+	}
+
+	return config.defaultPriority
+}
+
+// =============================================================================
+// Translation Functions
+// =============================================================================
+
+/**
+ * Translate a single TodoItem to a PlanEntry.
+ *
+ * @param item - The todo item to translate
+ * @param index - Position in the list (0-based)
+ * @param total - Total number of items
+ * @param config - Priority configuration
+ * @returns The translated plan entry
+ */
+export function todoItemToPlanEntry(
+	item: TodoItem,
+	index: number = 0,
+	total: number = 1,
+	config: PriorityConfig = DEFAULT_PRIORITY_CONFIG,
+): PlanEntry {
+	return {
+		content: item.content,
+		priority: determinePriority(item, index, total, config),
+		status: item.status,
+	}
+}
+
+/**
+ * Translate an array of TodoItems to a PlanUpdate.
+ *
+ * @param todos - Array of todo items
+ * @param config - Optional partial priority configuration
+ * @returns The plan update payload
+ */
+export function todoListToPlanUpdate(todos: TodoItem[], config?: Partial<PriorityConfig>): PlanUpdate {
+	const mergedConfig: PriorityConfig = { ...DEFAULT_PRIORITY_CONFIG, ...config }
+	const total = todos.length
+
+	return {
+		sessionUpdate: "plan",
+		entries: todos.map((item, index) => todoItemToPlanEntry(item, index, total, mergedConfig)),
+	}
+}
+
+// =============================================================================
+// Message Detection and Parsing
+// =============================================================================
+
+/**
+ * Parsed todo list message structure.
+ */
+interface ParsedTodoMessage {
+	tool: "updateTodoList"
+	todos: TodoItem[]
+}
+
+/**
+ * Type guard to check if parsed JSON is a valid todo list message.
+ */
+function isParsedTodoMessage(obj: unknown): obj is ParsedTodoMessage {
+	if (!obj || typeof obj !== "object") return false
+	const record = obj as Record<string, unknown>
+	return record.tool === "updateTodoList" && Array.isArray(record.todos)
+}
+
+/**
+ * Parse todo list from a tool message text.
+ *
+ * @param text - The message text (JSON string)
+ * @returns Array of TodoItems or null if not a valid todo message
+ */
+export function parseTodoListFromMessage(text: string): TodoItem[] | null {
+	try {
+		const parsed: unknown = JSON.parse(text)
+		if (isParsedTodoMessage(parsed)) {
+			return parsed.todos
+		}
+	} catch {
+		// Not valid JSON - ignore
+	}
+	return null
+}
+
+/**
+ * Minimal message interface for detection.
+ */
+interface MessageLike {
+	type: string
+	ask?: string
+	say?: string
+	text?: string
+}
+
+/**
+ * Check if a message contains a todo list update.
+ *
+ * Detects two types of messages:
+ * 1. Tool ask messages with updateTodoList
+ * 2. user_edit_todos say messages (when user edits the todo list)
+ *
+ * @param message - The message to check
+ * @returns true if message contains a todo list update
+ */
+export function isTodoListMessage(message: MessageLike): boolean {
+	// Check for tool ask message with updateTodoList
+	if (message.type === "ask" && message.ask === "tool" && message.text) {
+		const todos = parseTodoListFromMessage(message.text)
+		return todos !== null
+	}
+
+	// Check for user_edit_todos say message
+	if (message.type === "say" && message.say === "user_edit_todos" && message.text) {
+		const todos = parseTodoListFromMessage(message.text)
+		return todos !== null
+	}
+
+	return false
+}
+
+/**
+ * Extract todo list from a message if present.
+ *
+ * @param message - The message to extract from
+ * @returns Array of TodoItems or null if not a todo message
+ */
+export function extractTodoListFromMessage(message: MessageLike): TodoItem[] | null {
+	if (!message.text) return null
+
+	if (message.type === "ask" && message.ask === "tool") {
+		return parseTodoListFromMessage(message.text)
+	}
+
+	if (message.type === "say" && message.say === "user_edit_todos") {
+		return parseTodoListFromMessage(message.text)
+	}
+
+	return null
+}
+
+/**
+ * Create a plan update from a message if it contains a todo list.
+ *
+ * Convenience function that combines detection, extraction, and translation.
+ *
+ * @param message - The message to process
+ * @param config - Optional priority configuration
+ * @returns PlanUpdate or null if message doesn't contain todos
+ */
+export function createPlanUpdateFromMessage(message: MessageLike, config?: Partial<PriorityConfig>): PlanUpdate | null {
+	const todos = extractTodoListFromMessage(message)
+	if (!todos || todos.length === 0) {
+		return null
+	}
+	return todoListToPlanUpdate(todos, config)
+}

apps/cli/src/acp/translator/prompt-extractor.ts

@@ -0,0 +1,101 @@
+/**
+ * Prompt Extractor
+ *
+ * Extracts text and images from ACP prompt content blocks.
+ * Handles various content block types including text, resources, and media.
+ */
+
+import type * as acp from "@agentclientprotocol/sdk"
+
+// =============================================================================
+// Text Extraction
+// =============================================================================
+
+/**
+ * Extract text content from ACP prompt content blocks.
+ *
+ * Handles these content block types:
+ * - text: Direct text content
+ * - resource_link: Reference to a file or resource (converted to @uri format)
+ * - resource: Embedded resource with text content
+ * - image/audio: Noted as placeholders
+ *
+ * @param prompt - Array of ACP content blocks
+ * @returns Combined text from all blocks
+ */
+export function extractPromptText(prompt: acp.ContentBlock[]): string {
+	const textParts: string[] = []
+
+	for (const block of prompt) {
+		switch (block.type) {
+			case "text":
+				textParts.push(block.text)
+				break
+			case "resource_link":
+				// Reference to a file or resource
+				textParts.push(`@${block.uri}`)
+				break
+			case "resource":
+				// Embedded resource content
+				if (block.resource && "text" in block.resource) {
+					textParts.push(`Content from ${block.resource.uri}:\n${block.resource.text}`)
+				}
+				break
+			case "image":
+			case "audio":
+				// Binary content - note it but don't include
+				textParts.push(`[${block.type} content]`)
+				break
+		}
+	}
+
+	return textParts.join("\n")
+}
+
+// =============================================================================
+// Image Extraction
+// =============================================================================
+
+/**
+ * Extract images from ACP prompt content blocks.
+ *
+ * Extracts base64-encoded image data from image content blocks.
+ *
+ * @param prompt - Array of ACP content blocks
+ * @returns Array of base64-encoded image data strings
+ */
+export function extractPromptImages(prompt: acp.ContentBlock[]): string[] {
+	const images: string[] = []
+
+	for (const block of prompt) {
+		if (block.type === "image" && block.data) {
+			images.push(block.data)
+		}
+	}
+
+	return images
+}
+
+// =============================================================================
+// Resource Extraction
+// =============================================================================
+
+/**
+ * Extract resource URIs from ACP prompt content blocks.
+ *
+ * @param prompt - Array of ACP content blocks
+ * @returns Array of resource URIs
+ */
+export function extractPromptResources(prompt: acp.ContentBlock[]): string[] {
+	const resources: string[] = []
+
+	for (const block of prompt) {
+		if (block.type === "resource_link") {
+			resources.push(block.uri)
+		} else if (block.type === "resource" && block.resource) {
+			resources.push(block.resource.uri)
+		}
+	}
+
+	return resources
+}

apps/cli/src/acp/translator/tool-parser.ts

@@ -0,0 +1,241 @@
+/**
+ * Tool Parser
+ *
+ * Parses tool information from ClineMessage format.
+ * Extracts tool name, parameters, and generates titles.
+ */
+
+import * as path from "node:path"
+import type * as acp from "@agentclientprotocol/sdk"
+import type { ClineMessage } from "@roo-code/types"
+
+import { mapToolToKind, isEditTool as isFileEditTool } from "../tool-registry.js"
+import { extractLocations } from "./location-extractor.js"
+import { parseUnifiedDiff } from "./diff-parser.js"
+import { resolveFilePathUnsafe } from "../utils/index.js"
+
+// =============================================================================
+// Types
+// =============================================================================
+
+/**
+ * Parsed tool call information.
+ */
+export interface ToolCallInfo {
+	/** Unique identifier for the tool call */
+	id: string
+	/** Tool name */
+	name: string
+	/** Human-readable title */
+	title: string
+	/** Tool parameters */
+	params: Record<string, unknown>
+	/** File locations involved */
+	locations: acp.ToolCallLocation[]
+	/** Tool content (diffs, etc.) */
+	content?: acp.ToolCallContent[]
+}
+
+// =============================================================================
+// Tool Call ID Generation
+// =============================================================================
+
+/**
+ * Generate a tool call ID from a ClineMessage timestamp.
+ *
+ * Uses the message timestamp directly, which provides:
+ * - Deterministic IDs - same message always produces same ID
+ * - Natural deduplication - duplicate waitingForInput events use same ID
+ * - Easy debugging - can correlate ACP tool calls to ClineMessages
+ * - Sortable by creation time
+ *
+ * @param timestamp - ClineMessage timestamp (message.ts)
+ * @returns Tool call ID
+ */
+function generateToolCallId(timestamp: number): string {
+	return `tool-${timestamp}`
+}
+
+// =============================================================================
+// Tool Parsing
+// =============================================================================
+
+/**
+ * Parse tool information from a ClineMessage.
+ *
+ * Handles two formats:
+ * 1. JSON format: Message text is JSON with tool name and parameters
+ * 2. Text format: Tool name extracted from text like "Using/Executing/Running X"
+ *
+ * @param message - The ClineMessage to parse
+ * @param workspacePath - Optional workspace path to resolve relative paths
+ * @returns Parsed tool info or null if parsing fails
+ */
+export function parseToolFromMessage(message: ClineMessage, workspacePath?: string): ToolCallInfo | null {
+	if (!message.text) {
+		return null
+	}
+
+	// Tool messages typically have JSON content describing the tool
+	try {
+		// Try to parse as JSON first
+		if (message.text.startsWith("{")) {
+			const parsed = JSON.parse(message.text) as Record<string, unknown>
+			const toolName = (parsed.tool as string) || "unknown"
+			const filePath = (parsed.path as string) || undefined
+
+			return {
+				id: generateToolCallId(message.ts),
+				name: toolName,
+				title: generateToolTitle(toolName, filePath),
+				params: parsed,
+				locations: extractLocations(parsed, workspacePath),
+				content: extractToolContent(parsed, workspacePath),
+			}
+		}
+	} catch {
+		// Not JSON, try to extract tool info from text
+	}
+
+	// Extract tool name from text content
+	const toolMatch = message.text.match(/(?:Using|Executing|Running)\s+(\w+)/i)
+	const toolName = toolMatch?.[1] || "unknown"
+
+	return {
+		id: generateToolCallId(message.ts),
+		name: toolName,
+		title: message.text.slice(0, 100),
+		params: {},
+		locations: [],
+	}
+}
+
+// =============================================================================
+// Tool Title Generation
+// =============================================================================
+
+/**
+ * Generate a human-readable title for a tool operation.
+ *
+ * Maps tool names to descriptive titles, optionally including file names.
+ *
+ * @param toolName - The tool name
+ * @param filePath - Optional file path for context
+ * @returns Human-readable title
+ */
+export function generateToolTitle(toolName: string, filePath?: string): string {
+	const fileName = filePath ? path.basename(filePath) : undefined
+
+	// Map tool names to human-readable titles
+	const toolTitles: Record<string, string> = {
+		// File creation
+		newFileCreated: fileName ? `Creating ${fileName}` : "Creating file",
+		write_to_file: fileName ? `Writing ${fileName}` : "Writing file",
+		create_file: fileName ? `Creating ${fileName}` : "Creating file",
+
+		// File editing
+		editedExistingFile: fileName ? `Edit ${fileName}` : "Edit file",
+		apply_diff: fileName ? `Edit ${fileName}` : "Edit file",
+		appliedDiff: fileName ? `Edit ${fileName}` : "Edit file",
+		modify_file: fileName ? `Edit ${fileName}` : "Edit file",
+
+		// File reading
+		read_file: fileName ? `Read ${fileName}` : "Read file",
+		readFile: fileName ? `Read ${fileName}` : "Read file",
+
+		// File listing
+		list_files: filePath ? `Listing files in ${filePath}` : "Listing files",
+		listFiles: filePath ? `Listing files in ${filePath}` : "Listing files",
+
+		// File search
+		search_files: "Searching files",
+		searchFiles: "Searching files",
+
+		// Command execution
+		execute_command: "Running command",
+		executeCommand: "Running command",
+
+		// Browser actions
+		browser_action: "Browser action",
+		browserAction: "Browser action",
+
+		// Plan updates
+		updateTodoList: "Update plan",
+		update_todo_list: "Update plan",
+	}
+
+	return toolTitles[toolName] || (fileName ? `${toolName}: ${fileName}` : toolName)
+}
+
+// =============================================================================
+// Tool Content Extraction
+// =============================================================================
+
+/**
+ * Extract tool content for ACP (diffs, text, etc.)
+ *
+ * For file edit tools, parses the content as a unified diff.
+ *
+ * @param params - Tool parameters
+ * @param workspacePath - Optional workspace path
+ * @returns Array of tool content or undefined
+ */
+export function extractToolContent(
+	params: Record<string, unknown>,
+	workspacePath?: string,
+): acp.ToolCallContent[] | undefined {
+	const content: acp.ToolCallContent[] = []
+
+	// Check if this is a file operation with diff content
+	const filePath = params.path as string | undefined
+	const diffContent = params.content as string | undefined
+	const toolName = params.tool as string | undefined
+
+	if (filePath && diffContent && isFileEditTool(toolName || "")) {
+		const absolutePath = resolveFilePathUnsafe(filePath, workspacePath)
+		const parsedDiff = parseUnifiedDiff(diffContent)
+
+		if (parsedDiff) {
+			// Use ACP diff format
+			content.push({
+				type: "diff",
+				path: absolutePath,
+				oldText: parsedDiff.oldText,
+				newText: parsedDiff.newText,
+			} as acp.ToolCallContent)
+		}
+	}
+
+	return content.length > 0 ? content : undefined
+}
+
+// =============================================================================
+// Tool Call Building
+// =============================================================================
+
+/**
+ * Build an ACP ToolCall from a ClineMessage.
+ *
+ * @param message - The ClineMessage to parse
+ * @param workspacePath - Optional workspace path to resolve relative paths
+ * @returns ACP ToolCall object
+ */
+export function buildToolCallFromMessage(message: ClineMessage, workspacePath?: string): acp.ToolCall {
+	const toolInfo = parseToolFromMessage(message, workspacePath)
+
+	const toolCall: acp.ToolCall = {
+		toolCallId: toolInfo?.id || generateToolCallId(message.ts),
+		title: toolInfo?.title || message.text?.slice(0, 100) || "Tool execution",
+		kind: toolInfo ? mapToolToKind(toolInfo.name) : "other",
+		status: "pending",
+		locations: toolInfo?.locations || [],
+		rawInput: toolInfo?.params || {},
+	}
+
+	// Include content if available (e.g., diffs for file operations)
+	if (toolInfo?.content && toolInfo.content.length > 0) {
+		toolCall.content = toolInfo.content
+	}
+
+	return toolCall
+}

apps/cli/src/acp/types.ts

@@ -0,0 +1,42 @@
+import type { ModelInfo, SessionMode } from "@agentclientprotocol/sdk"
+
+export const DEFAULT_MODELS: ModelInfo[] = [
+	{
+		modelId: "anthropic/claude-opus-4.5",
+		name: "Claude Opus 4.5",
+		description: "Most capable for complex work",
+	},
+	{
+		modelId: "anthropic/claude-sonnet-4.5",
+		name: "Claude Sonnet 4.5",
+		description: "Best balance of speed and capability",
+	},
+	{
+		modelId: "anthropic/claude-haiku-4.5",
+		name: "Claude Haiku 4.5",
+		description: "Fastest for quick answers",
+	},
+]
+
+export const AVAILABLE_MODES: SessionMode[] = [
+	{
+		id: "code",
+		name: "Code",
+		description: "Write, modify, and refactor code",
+	},
+	{
+		id: "architect",
+		name: "Architect",
+		description: "Plan and design system architecture",
+	},
+	{
+		id: "ask",
+		name: "Ask",
+		description: "Ask questions and get explanations",
+	},
+	{
+		id: "debug",
+		name: "Debug",
+		description: "Debug issues and troubleshoot problems",
+	},
+]

apps/cli/src/acp/utils/format-utils.ts

@@ -0,0 +1,379 @@
+/**
+ * Format Utilities
+ *
+ * Shared formatting and content extraction utilities for ACP.
+ * Extracted to eliminate code duplication across modules.
+ */
+
+import * as fs from "node:fs"
+import * as fsPromises from "node:fs/promises"
+import * as path from "node:path"
+
+// =============================================================================
+// Configuration
+// =============================================================================
+
+/**
+ * Default configuration for content formatting.
+ */
+export interface FormatConfig {
+	/** Maximum number of lines to show for read results */
+	maxReadLines: number
+}
+
+export const DEFAULT_FORMAT_CONFIG: FormatConfig = {
+	maxReadLines: 100,
+}
+
+// =============================================================================
+// Result Type for Error Handling
+// =============================================================================
+
+/**
+ * Result type for operations that can fail.
+ * Provides explicit success/failure indication instead of returning error strings.
+ */
+export type Result<T> = { ok: true; value: T } | { ok: false; error: string }
+
+/**
+ * Create a successful result.
+ */
+export function ok<T>(value: T): Result<T> {
+	return { ok: true, value }
+}
+
+/**
+ * Create a failed result.
+ */
+export function err<T>(error: string): Result<T> {
+	return { ok: false, error }
+}
+
+// =============================================================================
+// Search Result Formatting
+// =============================================================================
+
+/**
+ * Format search results into a clean summary with file list.
+ *
+ * Input format (verbose):
+ * ```
+ * Found 112 results.
+ *
+ * # src/acp/__tests__/agent.test.ts
+ *   9 |
+ *  10 | // Mock the auth module
+ * ...
+ *
+ * # README.md
+ * 105 |
+ * ...
+ * ```
+ *
+ * Output format (clean):
+ * ```
+ * Found 112 results in 20 files
+ *
+ * - src/acp/__tests__/agent.test.ts
+ * - README.md
+ * ...
+ * ```
+ */
+export function formatSearchResults(content: string): string {
+	// Extract count from "Found X results" line
+	const countMatch = content.match(/Found (\d+) results?/)
+	const resultCount = countMatch?.[1] ? parseInt(countMatch[1], 10) : null
+
+	// Extract unique file paths from "# path/to/file" lines
+	const filePattern = /^# (.+)$/gm
+	const files = new Set<string>()
+	let match
+	while ((match = filePattern.exec(content)) !== null) {
+		if (match[1]) {
+			files.add(match[1])
+		}
+	}
+
+	// Sort files alphabetically
+	const fileList = Array.from(files).sort((a, b) => a.localeCompare(b))
+
+	// Build the formatted output
+	if (fileList.length === 0) {
+		// No files found, return first line (might be "No results found" or similar)
+		return content.split("\n")[0] || content
+	}
+
+	const summary =
+		resultCount !== null
+			? `Found ${resultCount} result${resultCount !== 1 ? "s" : ""} in ${fileList.length} file${fileList.length !== 1 ? "s" : ""}`
+			: `Found matches in ${fileList.length} file${fileList.length !== 1 ? "s" : ""}`
+
+	// Use markdown list format
+	const formattedFiles = fileList.map((f) => `- ${f}`).join("\n")
+
+	return `${summary}\n\n${formattedFiles}`
+}
+
+// =============================================================================
+// Read Content Formatting
+// =============================================================================
+
+/**
+ * Format read results by truncating long file contents.
+ *
+ * @param content - The raw file content
+ * @param config - Optional configuration overrides
+ * @returns Truncated content with indicator if truncated
+ */
+export function formatReadContent(content: string, config: FormatConfig = DEFAULT_FORMAT_CONFIG): string {
+	const lines = content.split("\n")
+
+	if (lines.length <= config.maxReadLines) {
+		return content
+	}
+
+	// Truncate and add indicator
+	const truncated = lines.slice(0, config.maxReadLines).join("\n")
+	const remaining = lines.length - config.maxReadLines
+	return `${truncated}\n\n... (${remaining} more lines)`
+}
+
+// =============================================================================
+// Code Block Wrapping
+// =============================================================================
+
+/**
+ * Wrap content in markdown code block for better rendering.
+ *
+ * @param content - Content to wrap
+ * @param language - Optional language for syntax highlighting
+ * @returns Content wrapped in markdown code fences
+ */
+export function wrapInCodeBlock(content: string, language?: string): string {
+	const fence = language ? `\`\`\`${language}` : "```"
+	return `${fence}\n${content}\n\`\`\``
+}
+
+// =============================================================================
+// Content Extraction from Raw Input
+// =============================================================================
+
+/**
+ * Common field names to check when extracting content from tool parameters.
+ */
+const CONTENT_FIELDS = ["content", "text", "result", "output", "fileContent", "data"] as const
+
+/**
+ * Extract content from raw input parameters.
+ *
+ * Tries common field names for content. Returns the first non-empty string found.
+ *
+ * @param rawInput - Tool parameters object
+ * @returns Extracted content or undefined if not found
+ */
+export function extractContentFromParams(rawInput: Record<string, unknown>): string | undefined {
+	for (const field of CONTENT_FIELDS) {
+		const value = rawInput[field]
+		if (typeof value === "string" && value.length > 0) {
+			return value
+		}
+	}
+
+	return undefined
+}
+
+// =============================================================================
+// File Reading
+// =============================================================================
+
+/**
+ * Resolve a file path to absolute, using workspace path if relative.
+ * Includes path traversal protection when workspace path is provided.
+ *
+ * @param filePath - File path (may be relative or absolute)
+ * @param workspacePath - Workspace path for resolving relative paths
+ * @returns Result with absolute path, or error if path traversal detected
+ */
+export function resolveFilePath(filePath: string, workspacePath?: string): Result<string> {
+	// Normalize the path to resolve any . or .. segments
+	const normalizedPath = path.normalize(filePath)
+
+	if (path.isAbsolute(normalizedPath)) {
+		// For absolute paths with workspace, verify it's within workspace
+		if (workspacePath) {
+			const normalizedWorkspace = path.normalize(workspacePath)
+			if (!normalizedPath.startsWith(normalizedWorkspace + path.sep) && normalizedPath !== normalizedWorkspace) {
+				return err(`Path traversal detected: ${filePath} is outside workspace ${workspacePath}`)
+			}
+		}
+		return ok(normalizedPath)
+	}
+
+	if (workspacePath) {
+		const resolved = path.resolve(workspacePath, normalizedPath)
+		const normalizedWorkspace = path.normalize(workspacePath)
+
+		// Verify resolved path is within workspace (prevents ../../../etc/passwd attacks)
+		if (!resolved.startsWith(normalizedWorkspace + path.sep) && resolved !== normalizedWorkspace) {
+			return err(`Path traversal detected: ${filePath} resolves outside workspace ${workspacePath}`)
+		}
+
+		return ok(resolved)
+	}
+
+	// Return as-is if no workspace path available
+	return ok(normalizedPath)
+}
+
+/**
+ * Resolve a file path to absolute (legacy version without Result wrapper).
+ *
+ * @deprecated Use resolveFilePath() with Result type for better error handling
+ * @param filePath - File path (may be relative or absolute)
+ * @param workspacePath - Workspace path for resolving relative paths
+ * @returns Absolute path (returns original path on error)
+ */
+export function resolveFilePathUnsafe(filePath: string, workspacePath?: string): string {
+	const result = resolveFilePath(filePath, workspacePath)
+	return result.ok ? result.value : filePath
+}
+
+/**
+ * Read file content from the filesystem (synchronous version).
+ *
+ * For readFile tools, the rawInput.content field contains the file PATH
+ * (not the contents), so we need to read the actual file.
+ *
+ * @deprecated Use readFileContentAsync() for non-blocking I/O
+ * @param rawInput - Tool parameters (must contain path or content with file path)
+ * @param workspacePath - Workspace path for resolving relative paths
+ * @returns Result with file content or error message
+ */
+export function readFileContent(rawInput: Record<string, unknown>, workspacePath: string): Result<string> {
+	// The "content" field in readFile contains the absolute path
+	const filePath = rawInput.content as string | undefined
+	const relativePath = rawInput.path as string | undefined
+
+	// Try absolute path first, then relative path
+	let pathToRead: string | undefined
+	if (filePath) {
+		const resolved = resolveFilePath(filePath, workspacePath)
+		if (!resolved.ok) return resolved
+		pathToRead = resolved.value
+	} else if (relativePath) {
+		const resolved = resolveFilePath(relativePath, workspacePath)
+		if (!resolved.ok) return resolved
+		pathToRead = resolved.value
+	}
+
+	if (!pathToRead) {
+		return err("readFile tool has no path")
+	}
+
+	try {
+		const content = fs.readFileSync(pathToRead, "utf-8")
+		return ok(content)
+	} catch (error) {
+		return err(`Failed to read file ${pathToRead}: ${error}`)
+	}
+}
+
+/**
+ * Read file content from the filesystem (asynchronous version).
+ *
+ * For readFile tools, the rawInput.content field contains the file PATH
+ * (not the contents), so we need to read the actual file.
+ *
+ * @param rawInput - Tool parameters (must contain path or content with file path)
+ * @param workspacePath - Workspace path for resolving relative paths
+ * @returns Promise resolving to Result with file content or error message
+ */
+export async function readFileContentAsync(
+	rawInput: Record<string, unknown>,
+	workspacePath: string,
+): Promise<Result<string>> {
+	// The "content" field in readFile contains the absolute path
+	const filePath = rawInput.content as string | undefined
+	const relativePath = rawInput.path as string | undefined
+
+	// Try absolute path first, then relative path
+	let pathToRead: string | undefined
+	if (filePath) {
+		const resolved = resolveFilePath(filePath, workspacePath)
+		if (!resolved.ok) return resolved
+		pathToRead = resolved.value
+	} else if (relativePath) {
+		const resolved = resolveFilePath(relativePath, workspacePath)
+		if (!resolved.ok) return resolved
+		pathToRead = resolved.value
+	}
+
+	if (!pathToRead) {
+		return err("readFile tool has no path")
+	}
+
+	try {
+		const content = await fsPromises.readFile(pathToRead, "utf-8")
+		return ok(content)
+	} catch (error) {
+		return err(`Failed to read file ${pathToRead}: ${error}`)
+	}
+}
+
+// =============================================================================
+// User Echo Detection
+// =============================================================================
+
+/**
+ * Check if a text message is an echo of the user's prompt.
+ *
+ * When the extension starts processing a task, it often sends a `text`
+ * message containing the user's input. Since the ACP client already
+ * displays the user's message, we should filter this out.
+ *
+ * Uses fuzzy matching to handle minor differences (whitespace, etc.).
+ *
+ * @param text - The text to check
+ * @param promptText - The original prompt text to compare against
+ * @returns true if the text appears to be an echo of the prompt
+ */
+export function isUserEcho(text: string, promptText: string | null): boolean {
+	if (!promptText) {
+		return false
+	}
+
+	// Normalize both strings for comparison
+	const normalizedPrompt = promptText.trim().toLowerCase()
+	const normalizedText = text.trim().toLowerCase()
+
+	// Exact match
+	if (normalizedText === normalizedPrompt) {
+		return true
+	}
+
+	// Check if text is contained in prompt (might be truncated)
+	if (normalizedPrompt.includes(normalizedText) && normalizedText.length > 10) {
+		return true
+	}
+
+	// Check if prompt is contained in text (might have wrapper)
+	if (normalizedText.includes(normalizedPrompt) && normalizedPrompt.length > 10) {
+		return true
+	}
+
+	return false
+}
+
+// =============================================================================
+// Validation Helpers
+// =============================================================================
+
+/**
+ * Check if a path looks like a valid file path (has extension).
+ *
+ * @param filePath - Path to check
+ * @returns true if the path has a file extension
+ */
+export function hasValidFilePath(filePath: string): boolean {
+	return /\.[a-zA-Z0-9]+$/.test(filePath)
+}

apps/cli/src/acp/utils/index.ts

@@ -0,0 +1,29 @@
+/**
+ * ACP Utilities Module
+ *
+ * Shared utilities for the ACP implementation.
+ */
+
+export {
+	// Configuration
+	type FormatConfig,
+	DEFAULT_FORMAT_CONFIG,
+	// Result type
+	type Result,
+	ok,
+	err,
+	// Formatting functions
+	formatSearchResults,
+	formatReadContent,
+	wrapInCodeBlock,
+	// Content extraction
+	extractContentFromParams,
+	// File operations
+	readFileContent,
+	readFileContentAsync,
+	resolveFilePath,
+	resolveFilePathUnsafe,
+	// Validation
+	isUserEcho,
+	hasValidFilePath,
+} from "./format-utils.js"

apps/cli/src/agent/__tests__/extension-host.test.ts

@@ -98,12 +98,11 @@ describe("ExtensionHost", () => {
 
 			const host = new ExtensionHost(options)
 
-			// Options are stored but integrationTest is set to true
+			// Options are stored as provided
 			const storedOptions = getPrivate<ExtensionHostOptions>(host, "options")
 			expect(storedOptions.mode).toBe(options.mode)
 			expect(storedOptions.workspacePath).toBe(options.workspacePath)
 			expect(storedOptions.extensionPath).toBe(options.extensionPath)
-			expect(storedOptions.integrationTest).toBe(true) // Always set to true in constructor
 		})
 
 		it("should be an EventEmitter instance", () => {
@@ -281,8 +280,8 @@ describe("ExtensionHost", () => {
 	describe("quiet mode", () => {
 		describe("setupQuietMode", () => {
 			it("should not modify console when integrationTest is true", () => {
-				// By default, constructor sets integrationTest = true
-				const host = createTestHost()
+				// Explicitly set integrationTest = true
+				const host = createTestHost({ integrationTest: true })
 				const originalLog = console.log
 
 				callPrivate(host, "setupQuietMode")

apps/cli/src/agent/ask-dispatcher.ts

@@ -135,6 +135,7 @@ export class AskDispatcher {
 		}
 
 		// Skip partial messages (wait for complete)
+		// Note: Streaming output for partial tool/command messages is handled by OutputManager
 		if (message.partial) {
 			return { handled: false }
 		}
@@ -356,9 +357,12 @@ export class AskDispatcher {
 	 * Handle command execution approval.
 	 */
 	private async handleCommandApproval(ts: number, text: string): Promise<AskHandleResult> {
-		this.outputManager.output("\n[command request]")
-		this.outputManager.output(`  Command: ${text || "(no command specified)"}`)
-		this.outputManager.markDisplayed(ts, text || "", false)
+		// Skip output if we already streamed this command via partial messages
+		if (!this.outputManager.isAlreadyDisplayed(ts)) {
+			this.outputManager.output("\n[command request]")
+			this.outputManager.output(`  Command: ${text || "(no command specified)"}`)
+			this.outputManager.markDisplayed(ts, text || "", false)
+		}
 
 		if (this.nonInteractive) {
 			// Auto-approved by extension settings
@@ -380,46 +384,49 @@ export class AskDispatcher {
 	 * Handle tool execution approval.
 	 */
 	private async handleToolApproval(ts: number, text: string): Promise<AskHandleResult> {
-		let toolName = "unknown"
-		let toolInfo: Record<string, unknown> = {}
+		// Skip output if we already streamed this tool request via partial messages
+		if (!this.outputManager.isAlreadyDisplayed(ts)) {
+			let toolName = "unknown"
+			let toolInfo: Record<string, unknown> = {}
 
-		try {
-			toolInfo = JSON.parse(text) as Record<string, unknown>
-			toolName = (toolInfo.tool as string) || "unknown"
-		} catch {
-			// Use raw text if not JSON
-		}
-
-		const isProtected = toolInfo.isProtected === true
-
-		if (isProtected) {
-			this.outputManager.output(`\n[Tool Request] ${toolName} [PROTECTED CONFIGURATION FILE]`)
-			this.outputManager.output(`⚠️  WARNING: This tool wants to modify a protected configuration file.`)
-			this.outputManager.output(
-				`    Protected files include .rooignore, .roo/*, and other sensitive config files.`,
-			)
-		} else {
-			this.outputManager.output(`\n[Tool Request] ${toolName}`)
-		}
-
-		// Display tool details
-		for (const [key, value] of Object.entries(toolInfo)) {
-			if (key === "tool" || key === "isProtected") continue
-
-			let displayValue: string
-			if (typeof value === "string") {
-				displayValue = value.length > 200 ? value.substring(0, 200) + "..." : value
-			} else if (typeof value === "object" && value !== null) {
-				const json = JSON.stringify(value)
-				displayValue = json.length > 200 ? json.substring(0, 200) + "..." : json
-			} else {
-				displayValue = String(value)
+			try {
+				toolInfo = JSON.parse(text) as Record<string, unknown>
+				toolName = (toolInfo.tool as string) || "unknown"
+			} catch {
+				// Use raw text if not JSON
 			}
 
-			this.outputManager.output(`  ${key}: ${displayValue}`)
-		}
+			const isProtected = toolInfo.isProtected === true
 
-		this.outputManager.markDisplayed(ts, text || "", false)
+			if (isProtected) {
+				this.outputManager.output(`\n[Tool Request] ${toolName} [PROTECTED CONFIGURATION FILE]`)
+				this.outputManager.output(`⚠️  WARNING: This tool wants to modify a protected configuration file.`)
+				this.outputManager.output(
+					`    Protected files include .rooignore, .roo/*, and other sensitive config files.`,
+				)
+			} else {
+				this.outputManager.output(`\n[Tool Request] ${toolName}`)
+			}
+
+			// Display tool details
+			for (const [key, value] of Object.entries(toolInfo)) {
+				if (key === "tool" || key === "isProtected") continue
+
+				let displayValue: string
+				if (typeof value === "string") {
+					displayValue = value.length > 200 ? value.substring(0, 200) + "..." : value
+				} else if (typeof value === "object" && value !== null) {
+					const json = JSON.stringify(value)
+					displayValue = json.length > 200 ? json.substring(0, 200) + "..." : json
+				} else {
+					displayValue = String(value)
+				}
+
+				this.outputManager.output(`  ${key}: ${displayValue}`)
+			}
+
+			this.outputManager.markDisplayed(ts, text || "", false)
+		}
 
 		if (this.nonInteractive) {
 			// Auto-approved by extension settings (unless protected)

apps/cli/src/agent/events.ts

@@ -76,6 +76,11 @@ export interface ClientEventMap {
 	 */
 	modeChanged: ModeChangedEvent
 
+	/**
+	 * Emitted when command execution output is received (streaming terminal output).
+	 */
+	commandExecutionOutput: CommandExecutionOutputEvent
+
 	/**
 	 * Emitted on any error during message processing.
 	 */
@@ -128,6 +133,16 @@ export interface ModeChangedEvent {
 	currentMode: string
 }
 
+/**
+ * Event payload for command execution output (streaming terminal output).
+ */
+export interface CommandExecutionOutputEvent {
+	/** Unique execution ID */
+	executionId: string
+	/** The terminal output received so far */
+	output: string
+}
+
 // =============================================================================
 // Typed Event Emitter
 // =============================================================================

apps/cli/src/agent/extension-client.ts

@@ -432,6 +432,7 @@ export class ExtensionClient {
 		const message: WebviewMessage = {
 			type: "cancelTask",
 		}
+
 		this.sendMessage(message)
 	}
 
@@ -467,6 +468,7 @@ export class ExtensionClient {
 			type: "terminalOperation",
 			terminalOperation: "continue",
 		}
+
 		this.sendMessage(message)
 	}
 
@@ -480,6 +482,7 @@ export class ExtensionClient {
 			type: "terminalOperation",
 			terminalOperation: "abort",
 		}
+
 		this.sendMessage(message)
 	}
 

apps/cli/src/agent/extension-host.ts

@@ -31,12 +31,13 @@ import type { User } from "@/lib/sdk/index.js"
 import { getProviderSettings } from "@/lib/utils/provider.js"
 import { createEphemeralStorageDir } from "@/lib/storage/index.js"
 
-import type { WaitingForInputEvent, TaskCompletedEvent } from "./events.js"
+import type { WaitingForInputEvent, TaskCompletedEvent, CommandExecutionOutputEvent } from "./events.js"
 import type { AgentStateInfo } from "./agent-state.js"
 import { ExtensionClient } from "./extension-client.js"
 import { OutputManager } from "./output-manager.js"
 import { PromptManager } from "./prompt-manager.js"
 import { AskDispatcher } from "./ask-dispatcher.js"
+import { testLog } from "./test-logger.js"
 
 // Pre-configured logger for CLI message activity debugging.
 const cliLogger = new DebugLogger("CLI")
@@ -152,7 +153,7 @@ export class ExtensionHost extends EventEmitter implements ExtensionHostInterfac
 		super()
 
 		this.options = options
-		this.options.integrationTest = true
+		// this.options.integrationTest = true
 
 		// Initialize client - single source of truth for agent state (including mode).
 		this.client = new ExtensionClient({
@@ -163,6 +164,7 @@ export class ExtensionHost extends EventEmitter implements ExtensionHostInterfac
 		// Initialize output manager.
 		this.outputManager = new OutputManager({
 			disabled: options.disableOutput,
+			debug: options.debug,
 		})
 
 		// Initialize prompt manager with console mode callbacks.
@@ -189,6 +191,18 @@ export class ExtensionHost extends EventEmitter implements ExtensionHostInterfac
 			commandExecutionTimeout: 30,
 			browserToolEnabled: false,
 			enableCheckpoints: false,
+			// Disable preventFocusDisruption experiment for CLI - it's only
+			// relevant for VSCode diff views and preventing it causes tool
+			// messages to not stream during LLM generation.
+			experiments: {
+				multiFileApplyDiff: false,
+				powerSteering: false,
+				preventFocusDisruption: false,
+				imageGeneration: false,
+				runSlashCommand: false,
+				multipleNativeToolCalls: false,
+				customTools: false,
+			},
 			...getProviderSettings(this.options.provider, this.options.apiKey, this.options.model),
 		}
 
@@ -234,15 +248,53 @@ export class ExtensionHost extends EventEmitter implements ExtensionHostInterfac
 	 * The client emits events, managers handle them.
 	 */
 	private setupClientEventHandlers(): void {
+		// === TEST LOGGING: State changes (matches ACP session.ts logging) ===
+		this.client.on("stateChange", (event) => {
+			const prev = event.previousState
+			const curr = event.currentState
+
+			// Only log if something actually changed
+			const stateChanged =
+				prev.state !== curr.state ||
+				prev.isRunning !== curr.isRunning ||
+				prev.isStreaming !== curr.isStreaming ||
+				prev.currentAsk !== curr.currentAsk
+
+			if (stateChanged) {
+				testLog.info(
+					"ExtensionClient",
+					`STATE: ${prev.state} → ${curr.state} (running=${curr.isRunning}, streaming=${curr.isStreaming}, ask=${curr.currentAsk || "none"})`,
+				)
+			}
+		})
+
 		// Handle new messages - delegate to OutputManager.
 		this.client.on("message", (msg: ClineMessage) => {
 			this.logMessageDebug(msg, "new")
+
+			// DEBUG: Log all incoming messages with timestamp (only when -d flag is set)
+			if (this.options.debug) {
+				const ts = new Date().toISOString()
+				const msgType = msg.type === "say" ? `say:${msg.say}` : `ask:${msg.ask}`
+				const partial = msg.partial ? "PARTIAL" : "COMPLETE"
+				process.stdout.write(`\n[DEBUG ${ts}] NEW ${msgType} ${partial} ts=${msg.ts}\n`)
+			}
+
 			this.outputManager.outputMessage(msg)
 		})
 
 		// Handle message updates - delegate to OutputManager.
 		this.client.on("messageUpdated", (msg: ClineMessage) => {
 			this.logMessageDebug(msg, "updated")
+
+			// DEBUG: Log all message updates with timestamp (only when -d flag is set)
+			if (this.options.debug) {
+				const ts = new Date().toISOString()
+				const msgType = msg.type === "say" ? `say:${msg.say}` : `ask:${msg.ask}`
+				const partial = msg.partial ? "PARTIAL" : "COMPLETE"
+				process.stdout.write(`\n[DEBUG ${ts}] UPDATED ${msgType} ${partial} ts=${msg.ts}\n`)
+			}
+
 			this.outputManager.outputMessage(msg)
 		})
 
@@ -259,6 +311,11 @@ export class ExtensionHost extends EventEmitter implements ExtensionHostInterfac
 				this.outputManager.outputCompletionResult(event.message.ts, event.message.text || "")
 			}
 		})
+
+		// Handle streaming terminal output from commandExecutionStatus messages.
+		this.client.on("commandExecutionOutput", (event: CommandExecutionOutputEvent) => {
+			this.outputManager.outputStreamingTerminalOutput(event.executionId, event.output)
+		})
 	}
 
 	// ==========================================================================
@@ -436,9 +493,6 @@ export class ExtensionHost extends EventEmitter implements ExtensionHostInterfac
 		this.sendToExtension({ type: "newTask", text: prompt })
 
 		return new Promise((resolve, reject) => {
-			let timeoutId: NodeJS.Timeout | null = null
-			const timeoutMs: number = 110_000
-
 			const completeHandler = () => {
 				cleanup()
 				resolve()
@@ -450,23 +504,10 @@ export class ExtensionHost extends EventEmitter implements ExtensionHostInterfac
 			}
 
 			const cleanup = () => {
-				if (timeoutId) {
-					clearTimeout(timeoutId)
-					timeoutId = null
-				}
-
 				this.client.off("taskCompleted", completeHandler)
 				this.client.off("error", errorHandler)
 			}
 
-			// Set timeout to prevent indefinite hanging.
-			timeoutId = setTimeout(() => {
-				cleanup()
-				reject(
-					new Error(`Task completion timeout after ${timeoutMs}ms - no completion or error event received`),
-				)
-			}, timeoutMs)
-
 			this.client.once("taskCompleted", completeHandler)
 			this.client.once("error", errorHandler)
 		})

apps/cli/src/agent/index.ts

@@ -1 +1,3 @@
 export * from "./extension-host.js"
+export { ExtensionClient } from "./extension-client.js"
+export type { WaitingForInputEvent, TaskCompletedEvent, CommandExecutionOutputEvent } from "./events.js"

apps/cli/src/agent/message-processor.ts

@@ -21,7 +21,13 @@ import { ExtensionMessage, ClineMessage } from "@roo-code/types"
 import { debugLog } from "@roo-code/core/cli"
 
 import type { StateStore } from "./state-store.js"
-import type { TypedEventEmitter, AgentStateChangeEvent, WaitingForInputEvent, TaskCompletedEvent } from "./events.js"
+import type {
+	TypedEventEmitter,
+	AgentStateChangeEvent,
+	WaitingForInputEvent,
+	TaskCompletedEvent,
+	CommandExecutionOutputEvent,
+} from "./events.js"
 import {
 	isSignificantStateChange,
 	transitionedToWaiting,
@@ -121,6 +127,10 @@ export class MessageProcessor {
 					this.handleInvoke(message)
 					break
 
+				case "commandExecutionStatus":
+					this.handleCommandExecutionStatus(message)
+					break
+
 				default:
 					// Other message types are not relevant to state detection
 					if (this.options.debug) {
@@ -237,6 +247,7 @@ export class MessageProcessor {
 		}
 
 		const clineMessage = message.clineMessage
+
 		const previousState = this.store.getAgentState()
 
 		// Update the message in the store
@@ -277,6 +288,45 @@ export class MessageProcessor {
 		// But they might trigger state changes through subsequent messages
 	}
 
+	/**
+	 * Handle a "commandExecutionStatus" message - streaming terminal output.
+	 *
+	 * This message is sent during command execution to provide live terminal
+	 * output before the final command_output message is created.
+	 */
+	private handleCommandExecutionStatus(message: ExtensionMessage): void {
+		if (!message.text) {
+			return
+		}
+
+		try {
+			const status = JSON.parse(message.text) as { status: string; executionId?: string; output?: string }
+
+			// Only emit for "output" status which contains terminal output
+			if (status.status === "output" && status.executionId && status.output !== undefined) {
+				if (this.options.debug) {
+					debugLog("[MessageProcessor] Command execution output", {
+						executionId: status.executionId,
+						outputLength: status.output.length,
+					})
+				}
+
+				const event: CommandExecutionOutputEvent = {
+					executionId: status.executionId,
+					output: status.output,
+				}
+				this.emitter.emit("commandExecutionOutput", event)
+			}
+		} catch {
+			// Ignore parse errors
+			if (this.options.debug) {
+				debugLog("[MessageProcessor] Failed to parse commandExecutionStatus", {
+					text: message.text?.substring(0, 100),
+				})
+			}
+		}
+	}
+
 	// ===========================================================================
 	// Event Emission Helpers
 	// ===========================================================================
@@ -372,6 +422,15 @@ export class MessageProcessor {
 		// A more sophisticated implementation would track seen message timestamps
 		const lastMessage = messages[messages.length - 1]
 		if (lastMessage) {
+			// DEBUG: Log all emitted ask messages to trace partial handling
+			if (this.options.debug && lastMessage.type === "ask") {
+				debugLog("[MessageProcessor] EMIT message", {
+					ask: lastMessage.ask,
+					partial: lastMessage.partial,
+					textLen: lastMessage.text?.length || 0,
+					ts: lastMessage.ts,
+				})
+			}
 			this.emitter.emit("message", lastMessage)
 		}
 	}

apps/cli/src/agent/output-manager.ts

@@ -13,8 +13,17 @@
  * - Can be disabled for TUI mode where Ink controls the terminal
  */
 
+import fs from "fs"
 import { ClineMessage, ClineSay } from "@roo-code/types"
 
+// Debug logging to file (for CLI debugging without breaking TUI)
+const DEBUG_LOG = "/tmp/roo-cli-debug.log"
+function debugLog(message: string, data?: unknown) {
+	const timestamp = new Date().toISOString()
+	const entry = data ? `[${timestamp}] ${message}: ${JSON.stringify(data, null, 2)}\n` : `[${timestamp}] ${message}\n`
+	fs.appendFileSync(DEBUG_LOG, entry)
+}
+
 import { Observable } from "./events.js"
 
 // =============================================================================
@@ -58,6 +67,12 @@ export interface OutputManagerOptions {
 	 * Stream for error output (default: process.stderr).
 	 */
 	stderr?: NodeJS.WriteStream
+
+	/**
+	 * When true, outputs verbose debug info for tool requests.
+	 * Enabled by -d flag in CLI.
+	 */
+	debug?: boolean
 }
 
 // =============================================================================
@@ -68,6 +83,7 @@ export class OutputManager {
 	private disabled: boolean
 	private stdout: NodeJS.WriteStream
 	private stderr: NodeJS.WriteStream
+	private debug: boolean
 
 	/**
 	 * Track displayed messages by ts to avoid duplicate output.
@@ -90,6 +106,16 @@ export class OutputManager {
 	 */
 	private loggedFirstPartial = new Set<number>()
 
+	/**
+	 * Track streaming terminal output by execution ID.
+	 */
+	private terminalOutputByExecutionId = new Map<string, string>()
+
+	/**
+	 * Flag to track if we've streamed any terminal output (to skip command_output).
+	 */
+	private hasStreamedTerminalOutput = false
+
 	/**
 	 * Observable for streaming state changes.
 	 * External systems can subscribe to know when streaming starts/ends.
@@ -103,6 +129,7 @@ export class OutputManager {
 		this.disabled = options.disabled ?? false
 		this.stdout = options.stdout ?? process.stdout
 		this.stderr = options.stderr ?? process.stderr
+		this.debug = options.debug ?? false
 	}
 
 	// ===========================================================================
@@ -126,20 +153,48 @@ export class OutputManager {
 		if (msg.type === "say" && msg.say) {
 			this.outputSayMessage(ts, msg.say, text, isPartial, alreadyDisplayedComplete, skipFirstUserMessage)
 		} else if (msg.type === "ask" && msg.ask) {
-			// For ask messages, we only output command_output here
-			// Other asks are handled by AskDispatcher
-			if (msg.ask === "command_output") {
-				this.outputCommandOutput(ts, text, isPartial, alreadyDisplayedComplete)
+			// Handle streaming output for different ask types
+			switch (msg.ask) {
+				case "command_output":
+					this.outputCommandOutput(ts, text, isPartial, alreadyDisplayedComplete)
+					break
+
+				case "tool":
+					// Stream tool requests (file create/edit/delete) as they come in
+					this.outputToolRequest(ts, text, isPartial, alreadyDisplayedComplete)
+					break
+
+				case "command":
+					// Stream command requests as they come in
+					this.outputCommandRequest(ts, text, isPartial, alreadyDisplayedComplete)
+					break
+
+				// Other ask types (followup, completion_result, etc.) are handled by AskDispatcher
+				// when complete (partial: false)
 			}
 		}
 	}
 
+	/**
+	 * Get a timestamp for debug output.
+	 */
+	private getTimestamp(): string {
+		const now = new Date()
+		return `[${now.toISOString().slice(11, 23)}]`
+	}
+
+	/**
+	 * Whether to include timestamps in output (for debugging).
+	 */
+	private showTimestamps = !!process.env.DEBUG_TIMESTAMPS
+
 	/**
 	 * Output a simple text line with a label.
 	 */
 	output(label: string, text?: string): void {
 		if (this.disabled) return
-		const message = text ? `${label} ${text}\n` : `${label}\n`
+		const ts = this.showTimestamps ? `${this.getTimestamp()} ` : ""
+		const message = text ? `${ts}${label} ${text}\n` : `${ts}${label}\n`
 		this.stdout.write(message)
 	}
 
@@ -148,7 +203,8 @@ export class OutputManager {
 	 */
 	outputError(label: string, text?: string): void {
 		if (this.disabled) return
-		const message = text ? `${label} ${text}\n` : `${label}\n`
+		const ts = this.showTimestamps ? `${this.getTimestamp()} ` : ""
+		const message = text ? `${ts}${label} ${text}\n` : `${ts}${label}\n`
 		this.stderr.write(message)
 	}
 
@@ -157,13 +213,21 @@ export class OutputManager {
 	 */
 	writeRaw(text: string): void {
 		if (this.disabled) return
-		this.stdout.write(text)
+		const ts = this.showTimestamps ? `${this.getTimestamp()} ` : ""
+		this.stdout.write(ts + text)
 	}
 
 	/**
-	 * Check if a message has already been fully displayed.
+	 * Check if a message has already been displayed (streamed or complete).
+	 * Returns true if we've streamed content for this ts OR if we've fully displayed it.
 	 */
 	isAlreadyDisplayed(ts: number): boolean {
+		// Check if we've streamed any content for this message
+		// (streamedContent is set during streaming, before displayedMessages is finalized)
+		if (this.streamedContent.has(ts)) {
+			return true
+		}
+		// Check if we've fully displayed this message
 		const displayed = this.displayedMessages.get(ts)
 		return displayed !== undefined && !displayed.partial
 	}
@@ -198,6 +262,11 @@ export class OutputManager {
 		this.streamedContent.clear()
 		this.currentlyStreamingTs = null
 		this.loggedFirstPartial.clear()
+		this.terminalOutputByExecutionId.clear()
+		this.hasStreamedTerminalOutput = false
+		this.toolContentStreamed.clear()
+		this.toolContentTruncated.clear()
+		this.toolLastDisplayedCharCount.clear()
 		this.streamingState.next({ ts: null, isStreaming: false })
 	}
 
@@ -335,6 +404,7 @@ export class OutputManager {
 
 	/**
 	 * Output command_output (shared between say and ask types).
+	 * Skips output if we've already streamed terminal output via commandExecutionStatus.
 	 */
 	outputCommandOutput(
 		ts: number,
@@ -342,6 +412,15 @@ export class OutputManager {
 		isPartial: boolean,
 		alreadyDisplayedComplete: boolean | undefined,
 	): void {
+		// Skip if we've already streamed terminal output - avoid duplicate display
+		if (this.hasStreamedTerminalOutput) {
+			// Mark as displayed but don't output - we already showed it via [terminal]
+			if (!isPartial) {
+				this.displayedMessages.set(ts, { ts, text, partial: false })
+			}
+			return
+		}
+
 		if (isPartial && text) {
 			this.streamContent(ts, text, "[command output]")
 			this.displayedMessages.set(ts, { ts, text, partial: true })
@@ -365,6 +444,209 @@ export class OutputManager {
 		}
 	}
 
+	/**
+	 * Track streamed tool content separately (content grows, not the full JSON text).
+	 */
+	private toolContentStreamed = new Map<number, string>()
+
+	/**
+	 * Track which tool messages have already shown truncation marker.
+	 */
+	private toolContentTruncated = new Set<number>()
+
+	/**
+	 * Track the last displayed character count for streaming updates.
+	 */
+	private toolLastDisplayedCharCount = new Map<number, number>()
+
+	/**
+	 * Maximum lines to show when streaming file content.
+	 */
+	private static readonly MAX_PREVIEW_LINES = 5
+
+	/**
+	 * Helper to write debug output to stderr with timestamp.
+	 */
+	private debugOutput(message: string): void {
+		if (!this.debug) return
+		const ts = this.getTimestamp()
+		this.stderr.write(`${ts} [DEBUG] ${message}\n`)
+	}
+
+	/**
+	 * Output tool request (file create/edit/delete) with streaming content preview.
+	 * Shows the file content being written (up to 20 lines), then final state when complete.
+	 */
+	private outputToolRequest(
+		ts: number,
+		text: string,
+		isPartial: boolean,
+		alreadyDisplayedComplete: boolean | undefined,
+	): void {
+		// Parse tool info to get the tool name, path, and content for display
+		let toolName = "tool"
+		let toolPath = ""
+		let content = ""
+		try {
+			const toolInfo = JSON.parse(text) as Record<string, unknown>
+			toolName = (toolInfo.tool as string) || "tool"
+			toolPath = (toolInfo.path as string) || ""
+			content = (toolInfo.content as string) || ""
+		} catch {
+			// Use default if not JSON
+		}
+
+		// Debug output: show every tool request message
+		this.debugOutput(
+			`outputToolRequest: ts=${ts} partial=${isPartial} tool=${toolName} path="${toolPath}" contentLen=${content.length}`,
+		)
+
+		debugLog("[outputToolRequest] called", {
+			ts,
+			isPartial,
+			toolName,
+			toolPath,
+			contentLen: content.length,
+		})
+
+		if (isPartial && text) {
+			const previousContent = this.toolContentStreamed.get(ts) || ""
+			const previous = this.streamedContent.get(ts)
+			const currentLineCount = content === "" ? 0 : content.split("\n").length
+
+			// Check for valid extension: must have a dot followed by 1+ characters
+			const hasValidExtension = /\.[a-zA-Z0-9]+$/.test(toolPath)
+
+			// Don't show header until we have BOTH a valid path AND some content.
+			// This prevents showing "[newFileCreated] (0 chars)" followed by a long
+			// pause while the LLM generates the content.
+			const shouldShowHeader = hasValidExtension && content.length > 0
+
+			if (!previous && shouldShowHeader) {
+				// First partial with valid path and content - show header
+				const pathInfo = ` ${toolPath}`
+				debugLog("[outputToolRequest] FIRST PARTIAL - header", {
+					toolName,
+					toolPath,
+					contentLen: content.length,
+				})
+				this.writeRaw(`\n[${toolName}]${pathInfo} (${content.length} chars)\n`)
+				this.streamedContent.set(ts, { ts, text, headerShown: true })
+				this.toolLastDisplayedCharCount.set(ts, content.length)
+				this.currentlyStreamingTs = ts
+				this.streamingState.next({ ts, isStreaming: true })
+			} else if (!previous && !shouldShowHeader) {
+				// Early partial without valid path/content - track but don't show yet
+				// Just set headerShown: false to track we've seen this ts
+				this.streamedContent.set(ts, { ts, text, headerShown: false })
+			} else if (previous && !previous.headerShown && shouldShowHeader) {
+				// Path and content now valid - show the header now
+				const pathInfo = ` ${toolPath}`
+				debugLog("[outputToolRequest] DEFERRED HEADER", { toolName, toolPath, contentLen: content.length })
+				this.writeRaw(`\n[${toolName}]${pathInfo} (${content.length} chars)\n`)
+				this.streamedContent.set(ts, { ts, text, headerShown: true })
+				this.toolLastDisplayedCharCount.set(ts, content.length)
+				this.currentlyStreamingTs = ts
+				this.streamingState.next({ ts, isStreaming: true })
+			}
+
+			// Stream content delta (new content since last update)
+			if (content.length > previousContent.length && content.startsWith(previousContent)) {
+				const delta = content.slice(previousContent.length)
+				// Check if we're still within the preview limit
+				const previousLineCount = previousContent === "" ? 0 : previousContent.split("\n").length
+				const previouslyTruncated = this.toolContentTruncated.has(ts)
+
+				if (!previouslyTruncated) {
+					if (currentLineCount <= OutputManager.MAX_PREVIEW_LINES) {
+						// Still under limit - output the delta
+						this.writeRaw(delta)
+					} else if (previousLineCount < OutputManager.MAX_PREVIEW_LINES) {
+						// Just crossed the limit - output remaining lines up to limit, mark as truncated
+						const linesToShow = OutputManager.MAX_PREVIEW_LINES - previousLineCount
+						const deltaLines = delta.split("\n")
+						const truncatedDelta = deltaLines.slice(0, linesToShow).join("\n")
+						if (truncatedDelta) {
+							this.writeRaw(truncatedDelta)
+						}
+						this.toolContentTruncated.add(ts)
+						// Show streaming indicator with char count
+						this.writeRaw(`\n... streaming (${content.length} chars)`)
+						this.toolLastDisplayedCharCount.set(ts, content.length)
+					} else {
+						// Already at/past limit but not yet marked - just mark as truncated
+						this.toolContentTruncated.add(ts)
+					}
+				} else {
+					// Already truncated - update streaming char count on each update
+					// Output on new lines so updates are visible in captured output
+					const lastDisplayed = this.toolLastDisplayedCharCount.get(ts) || 0
+					if (content.length !== lastDisplayed) {
+						this.writeRaw(`\n... streaming (${content.length} chars)`)
+						this.toolLastDisplayedCharCount.set(ts, content.length)
+					}
+				}
+				this.toolContentStreamed.set(ts, content)
+			}
+
+			this.displayedMessages.set(ts, { ts, text, partial: true })
+		} else if (!isPartial && !alreadyDisplayedComplete) {
+			// Tool request complete
+			const previousContent = this.toolContentStreamed.get(ts) || ""
+			const currentLineCount = content === "" ? 0 : content.split("\n").length
+			const wasTruncated = this.toolContentTruncated.has(ts)
+
+			// Show final truncation message
+			if (wasTruncated && previousContent) {
+				const remainingLines = currentLineCount - OutputManager.MAX_PREVIEW_LINES
+				this.writeRaw(`\n... (${remainingLines} more lines)\n`)
+			}
+
+			// Show final stats
+			const pathInfo = toolPath ? ` ${toolPath}` : ""
+			const charCount = content.length
+			this.writeRaw(`[${toolName}]${pathInfo} complete (${currentLineCount} lines, ${charCount} chars)\n`)
+			this.currentlyStreamingTs = null
+			this.streamingState.next({ ts: null, isStreaming: false })
+			this.displayedMessages.set(ts, { ts, text, partial: false })
+			// Clean up tool content tracking
+			this.toolContentStreamed.delete(ts)
+			this.toolContentTruncated.delete(ts)
+			this.toolLastDisplayedCharCount.delete(ts)
+		}
+	}
+
+	/**
+	 * Output command request with streaming support.
+	 * Streams partial content as it arrives from the LLM.
+	 */
+	private outputCommandRequest(
+		ts: number,
+		text: string,
+		isPartial: boolean,
+		alreadyDisplayedComplete: boolean | undefined,
+	): void {
+		if (isPartial && text) {
+			this.streamContent(ts, text, "[command]")
+			this.displayedMessages.set(ts, { ts, text, partial: true })
+		} else if (!isPartial && !alreadyDisplayedComplete) {
+			// Command request complete - finish the stream
+			// Note: AskDispatcher will handle the actual prompt/approval
+			const streamed = this.streamedContent.get(ts)
+
+			if (streamed) {
+				if (text.length > streamed.text.length && text.startsWith(streamed.text)) {
+					const delta = text.slice(streamed.text.length)
+					this.writeRaw(delta)
+				}
+				this.finishStream(ts)
+			}
+			// Don't output non-streamed content here - AskDispatcher handles complete command requests
+
+			this.displayedMessages.set(ts, { ts, text, partial: false })
+		}
+	}
+
 	// ===========================================================================
 	// Streaming Helpers
 	// ===========================================================================
@@ -411,4 +693,38 @@ export class OutputManager {
 			this.displayedMessages.set(ts, { ts, text: text || "", partial: false })
 		}
 	}
+
+	// ===========================================================================
+	// Terminal Output Streaming (commandExecutionStatus)
+	// ===========================================================================
+
+	/**
+	 * Output streaming terminal output from commandExecutionStatus messages.
+	 * This provides live terminal output during command execution, before
+	 * the final command_output message is created.
+	 *
+	 * @param executionId - Unique execution ID for this command
+	 * @param output - The accumulated terminal output so far
+	 */
+	outputStreamingTerminalOutput(executionId: string, output: string): void {
+		if (this.disabled) return
+
+		// Mark that we've streamed terminal output (to skip command_output later)
+		this.hasStreamedTerminalOutput = true
+
+		const previousOutput = this.terminalOutputByExecutionId.get(executionId)
+
+		if (!previousOutput) {
+			// First time seeing this execution - output header and initial content
+			this.writeRaw("\n[terminal] ")
+			this.writeRaw(output)
+			this.terminalOutputByExecutionId.set(executionId, output)
+		} else if (output.length > previousOutput.length && output.startsWith(previousOutput)) {
+			// Output has grown - write only the delta
+			const delta = output.slice(previousOutput.length)
+			this.writeRaw(delta)
+			this.terminalOutputByExecutionId.set(executionId, output)
+		}
+		// If output hasn't grown or doesn't start with previous, ignore (likely reset)
+	}
 }

apps/cli/src/agent/test-logger.ts

@@ -0,0 +1,115 @@
+/**
+ * Test Logger for CLI/ACP Cancellation Debugging
+ *
+ * This writes logs to ~/.roo/cli-acp-test.log for comparing CLI
+ * behavior with ACP during cancellation testing.
+ *
+ * Format matches ACP logger for easy side-by-side comparison.
+ */
+
+import * as fs from "node:fs"
+import * as path from "node:path"
+import * as os from "node:os"
+
+const LOG_DIR = path.join(os.homedir(), ".roo")
+const LOG_FILE = path.join(LOG_DIR, "cli-acp-test.log")
+
+let stream: fs.WriteStream | null = null
+
+/**
+ * Ensure log file and directory exist.
+ */
+function ensureLogFile(): void {
+	try {
+		if (!fs.existsSync(LOG_DIR)) {
+			fs.mkdirSync(LOG_DIR, { recursive: true })
+		}
+		if (!stream) {
+			stream = fs.createWriteStream(LOG_FILE, { flags: "a" })
+		}
+	} catch {
+		// Silently fail
+	}
+}
+
+/**
+ * Format and write a log entry.
+ */
+function write(level: string, component: string, message: string, data?: unknown): void {
+	ensureLogFile()
+	if (!stream) return
+
+	const timestamp = new Date().toISOString()
+	let formatted = `[${timestamp}] [${level}] [${component}] ${message}`
+
+	if (data !== undefined) {
+		try {
+			const dataStr = JSON.stringify(data, null, 2)
+			formatted += `\n${dataStr}`
+		} catch {
+			formatted += ` [Data: unserializable]`
+		}
+	}
+
+	stream.write(formatted + "\n")
+}
+
+/**
+ * Test logger for CLI cancellation debugging.
+ *
+ * Usage:
+ *   testLog.info("ExtensionClient", "STATE: idle → running (running=true, streaming=true, ask=none)")
+ *   testLog.info("Session", "CANCEL: triggered")
+ */
+export const testLog = {
+	info(component: string, message: string, data?: unknown): void {
+		write("INFO", component, message, data)
+	},
+
+	debug(component: string, message: string, data?: unknown): void {
+		write("DEBUG", component, message, data)
+	},
+
+	warn(component: string, message: string, data?: unknown): void {
+		write("WARN", component, message, data)
+	},
+
+	error(component: string, message: string, data?: unknown): void {
+		write("ERROR", component, message, data)
+	},
+
+	/**
+	 * Clear the log file (call at start of test session).
+	 */
+	clear(): void {
+		try {
+			if (stream) {
+				stream.end()
+				stream = null
+			}
+			fs.writeFileSync(LOG_FILE, "")
+		} catch {
+			// Silently fail
+		}
+	},
+
+	/**
+	 * Get the log file path.
+	 */
+	getLogPath(): string {
+		return LOG_FILE
+	},
+
+	/**
+	 * Close the logger.
+	 */
+	close(): void {
+		if (stream) {
+			stream.end()
+			stream = null
+		}
+	},
+}
+
+// Log startup
+testLog.info("TestLogger", `CLI test logging initialized. Log file: ${LOG_FILE}`)

apps/cli/src/commands/acp/index.ts

@@ -0,0 +1,81 @@
+import { Readable, Writable } from "node:stream"
+import path from "node:path"
+import { fileURLToPath } from "node:url"
+
+import * as acpSdk from "@agentclientprotocol/sdk"
+
+import { type SupportedProvider, DEFAULT_FLAGS } from "@/types/index.js"
+import { getDefaultExtensionPath } from "@/lib/utils/extension.js"
+import { RooCodeAgent, acpLog } from "@/acp/index.js"
+
+export interface AcpCommandOptions {
+	extension?: string
+	provider?: SupportedProvider
+	model?: string
+	mode?: string
+	apiKey?: string
+}
+
+export async function runAcpServer(options: AcpCommandOptions): Promise<void> {
+	const __dirname = path.dirname(fileURLToPath(import.meta.url))
+	const extensionPath = options.extension || getDefaultExtensionPath(__dirname)
+
+	if (!extensionPath) {
+		console.error("Error: Extension path not found. Use --extension to specify the path.")
+		process.exit(1)
+	}
+
+	// Set up stdio streams for ACP communication.
+	// Note: We write to stdout (agent -> client) and read from stdin (client -> agent).
+	const stdout = Writable.toWeb(process.stdout) as WritableStream<Uint8Array>
+	const stdin = Readable.toWeb(process.stdin) as ReadableStream<Uint8Array>
+
+	const stream = acpSdk.ndJsonStream(stdout, stdin)
+	acpLog.info("Command", "ACP stream created, waiting for connection")
+
+	let agent: RooCodeAgent | null = null
+
+	const connection = new acpSdk.AgentSideConnection((conn: acpSdk.AgentSideConnection) => {
+		acpLog.info("Command", "Agent connection established")
+		agent = new RooCodeAgent(
+			{
+				extensionPath,
+				provider: options.provider ?? DEFAULT_FLAGS.provider,
+				model: options.model || DEFAULT_FLAGS.model,
+				mode: options.mode || DEFAULT_FLAGS.mode,
+				apiKey: options.apiKey || process.env.OPENROUTER_API_KEY,
+			},
+			conn,
+		)
+
+		return agent
+	}, stream)
+
+	const cleanup = async () => {
+		acpLog.info("Command", "Received shutdown signal, cleaning up")
+
+		if (agent) {
+			await agent.dispose()
+		}
+
+		acpLog.info("Command", "Cleanup complete, exiting")
+		process.exit(0)
+	}
+
+	process.on("SIGINT", cleanup)
+	process.on("SIGTERM", cleanup)
+
+	acpLog.info("Command", "Waiting for connection to close")
+	await connection.closed
+	acpLog.info("Command", "Connection closed")
+}
+
+export async function acp(options: AcpCommandOptions): Promise<void> {
+	try {
+		await runAcpServer(options)
+	} catch (error) {
+		acpLog.error("Command", "Fatal error", error)
+		console.error(error)
+		process.exit(1)
+	}
+}

apps/cli/src/commands/cli/run.ts

@@ -168,7 +168,7 @@ export async function run(workspaceArg: string, options: FlagOptions) {
 		console.log(ASCII_ROO)
 		console.log()
 		console.log(
-			`[roo] Running ${options.model || "default"} (${options.reasoningEffort || "default"}) on ${provider} in ${options.mode || "default"} mode in ${workspacePath}`,
+			`[roo] Running ${options.model || DEFAULT_FLAGS.model} (${options.reasoningEffort || "default"}) on ${provider} in ${options.mode || "default"} mode in ${workspacePath}`,
 		)
 
 		const host = new ExtensionHost({

apps/cli/src/commands/index.ts

@@ -1,2 +1,3 @@
 export * from "./auth/index.js"
 export * from "./cli/index.js"
+export * from "./acp/index.js"

apps/cli/src/index.ts

@@ -2,7 +2,7 @@ import { Command } from "commander"
 
 import { DEFAULT_FLAGS } from "@/types/constants.js"
 import { VERSION } from "@/lib/utils/version.js"
-import { run, login, logout, status } from "@/commands/index.js"
+import { run, login, logout, status, acp } from "@/commands/index.js"
 
 const program = new Command()
 
@@ -62,4 +62,14 @@ authCommand
 		process.exit(result.authenticated ? 0 : 1)
 	})
 
+program
+	.command("acp")
+	.description("Start ACP server mode for integration with editors like Zed")
+	.option("-e, --extension <path>", "Path to the extension bundle directory")
+	.option("-p, --provider <provider>", "API provider (anthropic, openai, openrouter, etc.)", DEFAULT_FLAGS.provider)
+	.option("-m, --model <model>", "Model to use", DEFAULT_FLAGS.model)
+	.option("-M, --mode <mode>", "Initial mode (code, architect, ask, debug)", DEFAULT_FLAGS.mode)
+	.option("-k, --api-key <key>", "API key for the LLM provider")
+	.action(acp)
+
 program.parse()

apps/cli/src/lib/utils/provider.ts

@@ -2,7 +2,7 @@ import { RooCodeSettings } from "@roo-code/types"
 
 import type { SupportedProvider } from "@/types/index.js"
 
-const envVarMap: Record<SupportedProvider, string> = {
+export const envVarMap: Record<SupportedProvider, string> = {
 	anthropic: "ANTHROPIC_API_KEY",
 	"openai-native": "OPENAI_API_KEY",
 	gemini: "GOOGLE_API_KEY",

apps/cli/src/types/constants.ts

@@ -4,6 +4,7 @@ export const DEFAULT_FLAGS = {
 	mode: "code",
 	reasoningEffort: "medium" as const,
 	model: "anthropic/claude-opus-4.5",
+	provider: "openrouter" as const,
 }
 
 export const REASONING_EFFORTS = [...reasoningEffortsExtended, "unspecified", "disabled"]

apps/cli/src/ui/App.tsx

@@ -18,12 +18,13 @@ import {
 	useTerminalSize,
 	useToast,
 	useExtensionHost,
-	useMessageHandlers,
 	useTaskSubmit,
 	useGlobalInput,
 	useFollowupCountdown,
 	useFocusManagement,
 	usePickerHandlers,
+	useClientEvents,
+	useExtensionState,
 } from "./hooks/index.js"
 
 // Import extracted utilities.
@@ -159,16 +160,14 @@ function AppInner({
 	// Toast notifications for ephemeral messages (e.g., mode changes).
 	const { currentToast, showInfo } = useToast()
 
-	const {
-		handleExtensionMessage,
-		seenMessageIds,
-		pendingCommandRef: _pendingCommandRef,
-		firstTextMessageSkipped,
-	} = useMessageHandlers({
-		nonInteractive,
-	})
+	// Handle non-message extension state (modes, file search, commands, task history)
+	const { handleExtensionState } = useExtensionState()
 
-	const { sendToExtension, runTask, cleanup } = useExtensionHost({
+	// Track seen message IDs and first text message skip for task submission
+	const seenMessageIds = useRef<Set<string>>(new Set())
+	const firstTextMessageSkipped = useRef(false)
+
+	const { client, sendToExtension, runTask, cleanup } = useExtensionHost({
 		initialPrompt,
 		mode,
 		reasoningEffort,
@@ -182,10 +181,27 @@ function AppInner({
 		nonInteractive,
 		ephemeral,
 		exitOnComplete,
-		onExtensionMessage: handleExtensionMessage,
+		onExtensionState: handleExtensionState,
 		createExtensionHost,
 	})
 
+	// Subscribe to ExtensionClient events for unified message handling
+	const { reset: resetClientEvents } = useClientEvents({
+		client,
+		nonInteractive,
+	})
+
+	// Reset tracking state when task is cleared
+	useEffect(() => {
+		if (!client) return
+		const unsubscribe = client.on("taskCleared" as "stateChange", () => {
+			seenMessageIds.current.clear()
+			firstTextMessageSkipped.current = false
+			resetClientEvents()
+		})
+		return unsubscribe
+	}, [client, resetClientEvents])
+
 	// Initialize task submit hook
 	const { handleSubmit, handleApprove, handleReject } = useTaskSubmit({
 		sendToExtension,

apps/cli/src/ui/components/ChatHistoryItem.tsx

@@ -1,5 +1,6 @@
-import { memo } from "react"
+// memo temporarily removed for debugging
 import { Box, Newline, Text } from "ink"
+import { DebugLogger } from "@roo-code/core/cli"
 
 import type { TUIMessage } from "../types.js"
 import * as theme from "../theme.js"
@@ -7,6 +8,8 @@ import * as theme from "../theme.js"
 import TodoDisplay from "./TodoDisplay.js"
 import { getToolRenderer } from "./tools/index.js"
 
+const renderLogger = new DebugLogger("RENDER")
+
 /**
  * Tool categories for styling
  */
@@ -215,6 +218,27 @@ function ChatHistoryItem({ message }: ChatHistoryItemProps) {
 				</Box>
 			)
 		case "tool": {
+			// Parse rawContent to get content for logging
+			let parsedContent = ""
+			try {
+				const parsed = JSON.parse(content) as Record<string, unknown>
+				parsedContent = ((parsed.content as string) || "").substring(0, 50)
+			} catch {
+				// Not JSON
+			}
+
+			renderLogger.debug("ChatHistoryItem:tool", {
+				id: message.id,
+				toolName: message.toolName,
+				hasToolData: !!message.toolData,
+				toolDataTool: message.toolData?.tool,
+				toolDataPath: message.toolData?.path,
+				toolDataContent: message.toolData?.content?.substring(0, 50),
+				rawContentLen: content.length,
+				parsedContent,
+				partial: message.partial,
+			})
+
 			// Special rendering for update_todo_list tool - show full TODO list
 			if (
 				(message.toolName === "update_todo_list" || message.toolName === "updateTodoList") &&
@@ -249,4 +273,5 @@ function ChatHistoryItem({ message }: ChatHistoryItemProps) {
 	}
 }
 
-export default memo(ChatHistoryItem)
+// Temporarily disable memo to debug streaming rendering issues
+export default ChatHistoryItem

apps/cli/src/ui/components/autocomplete/triggers/HelpTrigger.tsx

@@ -41,7 +41,6 @@ export function createHelpTrigger(): AutocompleteTrigger<HelpShortcutResult> {
 		id: "help",
 		triggerChar: "?",
 		position: "line-start",
-		consumeTrigger: true,
 
 		detectTrigger: (lineText: string): TriggerDetectionResult | null => {
 			// Check if line starts with ? (after optional whitespace)

apps/cli/src/ui/components/autocomplete/triggers/__tests__/HelpTrigger.test.tsx

@@ -157,13 +157,5 @@ describe("HelpTrigger", () => {
 			expect(trigger.emptyMessage).toBe("No matching shortcuts")
 			expect(trigger.debounceMs).toBe(0)
 		})
-
-		it("should have consumeTrigger set to true", () => {
-			const trigger = createHelpTrigger()
-
-			// The ? character should be consumed (not inserted into input)
-			// when the help menu is triggered
-			expect(trigger.consumeTrigger).toBe(true)
-		})
 	})
 })

apps/cli/src/ui/components/autocomplete/triggers/__tests__/HistoryTrigger.test.tsx

@@ -210,13 +210,6 @@ describe("HistoryTrigger", () => {
 			expect(trigger.debounceMs).toBe(100)
 		})
 
-		it("should not have consumeTrigger set (# character appears in input)", () => {
-			const trigger = createHistoryTrigger({ getHistory: () => mockHistoryItems })
-
-			// The # character should remain in the input like other triggers
-			expect(trigger.consumeTrigger).toBeUndefined()
-		})
-
 		it("should call getHistory when searching", () => {
 			const getHistoryMock = vi.fn(() => mockHistoryItems)
 			const trigger = createHistoryTrigger({ getHistory: getHistoryMock })

apps/cli/src/ui/components/autocomplete/types.ts

@@ -97,14 +97,6 @@ export interface AutocompleteTrigger<T extends AutocompleteItem = AutocompleteIt
 	 * @default 150
 	 */
 	debounceMs?: number
-
-	/**
-	 * Whether the trigger character should be consumed (not shown in input).
-	 * When true, the trigger character is treated as a control character
-	 * that activates the picker but doesn't appear in the text input.
-	 * @default false
-	 */
-	consumeTrigger?: boolean
 }
 
 /**

apps/cli/src/ui/components/autocomplete/useAutocompletePicker.ts

@@ -56,19 +56,6 @@ export function useAutocompletePicker<T extends AutocompleteItem>(
 		return lines[lines.length - 1] || ""
 	}, [])
 
-	/**
-	 * Get the input value with the trigger character removed.
-	 * Used when a trigger has consumeTrigger: true.
-	 */
-	const getConsumedValue = useCallback((value: string, lastLine: string, triggerIndex: number): string => {
-		const lines = value.split("\n")
-		const lastLineIndex = lines.length - 1
-		// Remove the trigger character from the last line
-		const newLastLine = lastLine.slice(0, triggerIndex) + lastLine.slice(triggerIndex + 1)
-		lines[lastLineIndex] = newLastLine
-		return lines.join("\n")
-	}, [])
-
 	/**
 	 * Handle input value changes - detects triggers and initiates search.
 	 * Returns an object indicating if the input should be modified (for consumeTrigger).
@@ -120,10 +107,6 @@ export function useAutocompletePicker<T extends AutocompleteItem>(
 
 			if (query === lastQuery && state.isOpen && state.activeTrigger?.id === foundTrigger.id) {
 				// Same query, same trigger - no need to search again
-				// Still return consumed value if trigger consumes input
-				if (foundTrigger.consumeTrigger) {
-					return { consumedValue: getConsumedValue(value, lastLine, foundTriggerInfo.triggerIndex) }
-				}
 				return {}
 			}
 
@@ -207,14 +190,9 @@ export function useAutocompletePicker<T extends AutocompleteItem>(
 
 			debounceTimersRef.current.set(foundTrigger.id, timer)
 
-			// Return consumed value if trigger consumes input
-			if (foundTrigger.consumeTrigger) {
-				return { consumedValue: getConsumedValue(value, lastLine, foundTriggerInfo.triggerIndex) }
-			}
-
 			return {}
 		},
-		[triggers, state.isOpen, state.activeTrigger?.id, getLastLine, getConsumedValue],
+		[triggers, state.isOpen, state.activeTrigger?.id, getLastLine],
 	)
 
 	/**

apps/cli/src/ui/components/tools/FileWriteTool.tsx

@@ -4,11 +4,11 @@ import * as theme from "../../theme.js"
 import { Icon } from "../Icon.js"
 
 import type { ToolRendererProps } from "./types.js"
-import { truncateText, sanitizeContent, getToolDisplayName, getToolIconName, parseDiff } from "./utils.js"
+import { sanitizeContent, getToolDisplayName, getToolIconName, parseDiff } from "./utils.js"
 
-const MAX_DIFF_LINES = 15
+const MAX_PREVIEW_LINES = 5
 
-export function FileWriteTool({ toolData }: ToolRendererProps) {
+export function FileWriteTool({ toolData, rawContent }: ToolRendererProps) {
 	const iconName = getToolIconName(toolData.tool)
 	const displayName = getToolDisplayName(toolData.tool)
 	const path = toolData.path || ""
@@ -18,6 +18,20 @@ export function FileWriteTool({ toolData }: ToolRendererProps) {
 	const isOutsideWorkspace = toolData.isOutsideWorkspace
 	const isNewFile = toolData.tool === "newFileCreated" || toolData.tool === "write_to_file"
 
+	// For streaming: rawContent is updated with each message, so parse it for live content
+	// toolData.content may be stale during streaming due to debounce optimization
+	let liveContent = toolData.content || ""
+	if (rawContent && isNewFile) {
+		try {
+			const parsed = JSON.parse(rawContent) as Record<string, unknown>
+			if (parsed.content && typeof parsed.content === "string") {
+				liveContent = parsed.content
+			}
+		} catch {
+			// Use toolData.content if rawContent isn't valid JSON
+		}
+	}
+
 	// Handle batch diff operations
 	if (toolData.batchDiffs && toolData.batchDiffs.length > 0) {
 		return (
@@ -57,9 +71,20 @@ export function FileWriteTool({ toolData }: ToolRendererProps) {
 	}
 
 	// Single file write
-	const { text: previewDiff, truncated, hiddenLines } = truncateText(diff, MAX_DIFF_LINES)
+	// For new files, display streaming content; for edits, show diff
 	const diffHunks = diff ? parseDiff(diff) : []
 
+	// Process content for display - split into lines and truncate
+	const sanitizedContent = isNewFile && liveContent ? sanitizeContent(liveContent) : ""
+	const contentLines = sanitizedContent ? sanitizedContent.split("\n") : []
+	const displayLines = contentLines.slice(0, MAX_PREVIEW_LINES)
+	const truncatedLineCount = contentLines.length - MAX_PREVIEW_LINES
+	const isContentTruncated = truncatedLineCount > 0
+
+	// Stats for the header
+	const totalLines = contentLines.length
+	const totalChars = liveContent.length
+
 	return (
 		<Box flexDirection="column" paddingX={1} marginBottom={1}>
 			{/* Header row with path on same line */}
@@ -76,15 +101,15 @@ export function FileWriteTool({ toolData }: ToolRendererProps) {
 						</Text>
 					</>
 				)}
-				{isNewFile && (
+				{isNewFile && !diffStats && (
 					<Text color={theme.successColor} bold>
 						{" "}
 						NEW
 					</Text>
 				)}
 
-				{/* Diff stats badge */}
-				{diffStats && (
+				{/* Stats - show line/char count for streaming, or diff stats when complete */}
+				{diffStats ? (
 					<>
 						<Text color={theme.dimText}> </Text>
 						<Text color={theme.successColor} bold>
@@ -95,6 +120,14 @@ export function FileWriteTool({ toolData }: ToolRendererProps) {
 							-{diffStats.removed}
 						</Text>
 					</>
+				) : (
+					isNewFile &&
+					totalChars > 0 && (
+						<Text color={theme.dimText}>
+							{" "}
+							({totalLines} lines, {totalChars} chars)
+						</Text>
+					)
 				)}
 
 				{/* Warning badges */}
@@ -107,7 +140,23 @@ export function FileWriteTool({ toolData }: ToolRendererProps) {
 				)}
 			</Box>
 
-			{/* Diff preview */}
+			{/* Streaming content preview for new files (before diff is available) */}
+			{isNewFile && !diff && displayLines.length > 0 && (
+				<Box flexDirection="column" marginLeft={2} marginTop={1}>
+					{displayLines.map((line, index) => (
+						<Text key={index} color={theme.toolText}>
+							{line}
+						</Text>
+					))}
+					{isContentTruncated && (
+						<Text color={theme.dimText} dimColor>
+							... ({truncatedLineCount} more lines)
+						</Text>
+					)}
+				</Box>
+			)}
+
+			{/* Diff preview for edits */}
 			{diffHunks.length > 0 && (
 				<Box flexDirection="column" marginLeft={2} marginTop={1}>
 					{diffHunks.slice(0, 2).map((hunk, hunkIndex) => (
@@ -149,13 +198,20 @@ export function FileWriteTool({ toolData }: ToolRendererProps) {
 				</Box>
 			)}
 
-			{/* Fallback to raw diff if no hunks parsed */}
-			{diffHunks.length === 0 && previewDiff && (
+			{/* Fallback: show raw diff content if no hunks parsed and not streaming new file */}
+			{!isNewFile && diffHunks.length === 0 && diff && (
 				<Box flexDirection="column" marginLeft={2} marginTop={1}>
-					<Text color={theme.toolText}>{previewDiff}</Text>
-					{truncated && (
+					{diff
+						.split("\n")
+						.slice(0, MAX_PREVIEW_LINES)
+						.map((line, index) => (
+							<Text key={index} color={theme.toolText}>
+								{line}
+							</Text>
+						))}
+					{diff.split("\n").length > MAX_PREVIEW_LINES && (
 						<Text color={theme.dimText} dimColor>
-							... ({hiddenLines} more lines)
+							... ({diff.split("\n").length - MAX_PREVIEW_LINES} more lines)
 						</Text>
 					)}
 				</Box>

apps/cli/src/ui/hooks/index.ts

@@ -6,17 +6,19 @@ export { useInputHistory } from "./useInputHistory.js"
 // Export new extracted hooks
 export { useFollowupCountdown } from "./useFollowupCountdown.js"
 export { useFocusManagement } from "./useFocusManagement.js"
-export { useMessageHandlers } from "./useMessageHandlers.js"
 export { useExtensionHost } from "./useExtensionHost.js"
 export { useTaskSubmit } from "./useTaskSubmit.js"
 export { useGlobalInput } from "./useGlobalInput.js"
 export { usePickerHandlers } from "./usePickerHandlers.js"
+export { useClientEvents } from "./useClientEvents.js"
+export { useExtensionState } from "./useExtensionState.js"
 
 // Export types
 export type { UseFollowupCountdownOptions } from "./useFollowupCountdown.js"
 export type { UseFocusManagementOptions, UseFocusManagementReturn } from "./useFocusManagement.js"
-export type { UseMessageHandlersOptions, UseMessageHandlersReturn } from "./useMessageHandlers.js"
 export type { UseExtensionHostOptions, UseExtensionHostReturn } from "./useExtensionHost.js"
 export type { UseTaskSubmitOptions, UseTaskSubmitReturn } from "./useTaskSubmit.js"
 export type { UseGlobalInputOptions } from "./useGlobalInput.js"
 export type { UsePickerHandlersOptions, UsePickerHandlersReturn } from "./usePickerHandlers.js"
+export type { UseClientEventsOptions, UseClientEventsReturn } from "./useClientEvents.js"
+export type { UseExtensionStateReturn } from "./useExtensionState.js"

apps/cli/src/ui/hooks/useClientEvents.ts

@@ -0,0 +1,516 @@
+/**
+ * useClientEvents - Bridge ExtensionClient events to TUI state
+ *
+ * This hook subscribes to ExtensionClient events (the same events used by
+ * non-TUI mode) and transforms them into TUI messages/state updates.
+ *
+ * This unifies the message handling logic between TUI and non-TUI modes:
+ * - Non-TUI: ExtensionClient events → OutputManager/AskDispatcher
+ * - TUI: ExtensionClient events → useClientEvents → Zustand store
+ */
+
+import { useEffect, useRef, useCallback } from "react"
+import type { ClineMessage, ClineAsk, ClineSay, TodoItem } from "@roo-code/types"
+import { consolidateTokenUsage, consolidateApiRequests, consolidateCommands, DebugLogger } from "@roo-code/core/cli"
+
+// Debug logger using same pattern as extension-host.ts
+const tuiLogger = new DebugLogger("TUI")
+
+import type { ExtensionClient } from "@/agent/index.js"
+import type { WaitingForInputEvent, CommandExecutionOutputEvent } from "@/agent/events.js"
+
+import type { TUIMessage, ToolData, PendingAsk } from "../types.js"
+import { useCLIStore } from "../store.js"
+import { extractToolData, formatToolOutput, formatToolAskMessage, parseTodosFromToolInfo } from "../utils/tools.js"
+
+export interface UseClientEventsOptions {
+	client: ExtensionClient | null
+	nonInteractive: boolean
+}
+
+export interface UseClientEventsReturn {
+	/** Reset tracking state (call when starting new task) */
+	reset: () => void
+}
+
+/**
+ * Hook that subscribes to ExtensionClient events and updates TUI state.
+ *
+ * Key events:
+ * - `message`: New ClineMessage → transform to TUIMessage and add to store
+ * - `messageUpdated`: Updated ClineMessage → update existing TUIMessage
+ * - `waitingForInput`: Ask needing input → set pendingAsk
+ */
+export function useClientEvents({ client, nonInteractive }: UseClientEventsOptions): UseClientEventsReturn {
+	const { addMessage, setPendingAsk, setLoading, setTokenUsage, currentTodos, setTodos } = useCLIStore()
+
+	// Track seen message timestamps to filter duplicates
+	const seenMessageIds = useRef<Set<string>>(new Set())
+	const firstTextMessageSkipped = useRef(false)
+
+	// Track pending command for injecting into command_output toolData
+	const pendingCommandRef = useRef<string | null>(null)
+
+	// Track the message ID of the current command being executed (for streaming updates)
+	const currentCommandMessageIdRef = useRef<string | null>(null)
+
+	// Track if we've streamed command output (to skip duplicate command_output say message)
+	const hasStreamedCommandOutputRef = useRef(false)
+
+	// Track the message ID of partial tool asks (for streaming file write updates)
+	const partialToolMessageIdRef = useRef<string | null>(null)
+
+	/**
+	 * Reset tracking state (call when starting new task)
+	 */
+	const reset = useCallback(() => {
+		seenMessageIds.current.clear()
+		firstTextMessageSkipped.current = false
+		pendingCommandRef.current = null
+		currentCommandMessageIdRef.current = null
+		hasStreamedCommandOutputRef.current = false
+		partialToolMessageIdRef.current = null
+	}, [])
+
+	/**
+	 * Transform a ClineMessage to TUIMessage and add to store
+	 */
+	const processClineMessage = useCallback(
+		(msg: ClineMessage) => {
+			const ts = msg.ts
+			const messageId = ts.toString()
+			const text = msg.text || ""
+			const partial = msg.partial || false
+			const isResuming = useCLIStore.getState().isResumingTask
+
+			// DEBUG: Log all ask messages to trace partial handling
+			if (msg.type === "ask") {
+				tuiLogger.debug("ask:received", {
+					ask: msg.ask,
+					partial,
+					textLen: text.length,
+					id: messageId,
+				})
+			}
+
+			if (msg.type === "say" && msg.say) {
+				processSayMessage(messageId, msg.say, text, partial, isResuming)
+			} else if (msg.type === "ask" && msg.ask) {
+				processAskMessage(messageId, msg.ask, text, partial)
+			}
+		},
+		[nonInteractive, currentTodos],
+	)
+
+	/**
+	 * Process "say" type messages
+	 */
+	const processSayMessage = useCallback(
+		(messageId: string, say: ClineSay, text: string, partial: boolean, isResuming: boolean) => {
+			// Skip certain message types
+			if (say === "checkpoint_saved" || say === "api_req_started" || say === "user_feedback") {
+				seenMessageIds.current.add(messageId)
+				return
+			}
+
+			// Skip first text message for new tasks (it's the user's prompt echo)
+			if (say === "text" && !firstTextMessageSkipped.current && !isResuming) {
+				firstTextMessageSkipped.current = true
+				seenMessageIds.current.add(messageId)
+				return
+			}
+
+			// Skip if already seen (non-partial)
+			if (seenMessageIds.current.has(messageId) && !partial) {
+				return
+			}
+
+			let role: TUIMessage["role"] = "assistant"
+			let toolName: string | undefined
+			let toolDisplayName: string | undefined
+			let toolDisplayOutput: string | undefined
+			let toolData: ToolData | undefined
+
+			if (say === "command_output") {
+				// Skip command_output say message if we've already streamed the output
+				// The streaming updates went to the command ask message directly
+				if (hasStreamedCommandOutputRef.current) {
+					seenMessageIds.current.add(messageId)
+					// Reset for next command
+					hasStreamedCommandOutputRef.current = false
+					currentCommandMessageIdRef.current = null
+					return
+				}
+
+				// Non-streamed case: add the command output message
+				role = "tool"
+				toolName = "execute_command"
+				toolDisplayName = "bash"
+				toolDisplayOutput = text
+				const trackedCommand = pendingCommandRef.current
+				toolData = { tool: "execute_command", command: trackedCommand || undefined, output: text }
+				pendingCommandRef.current = null
+			} else if (say === "reasoning") {
+				role = "thinking"
+			}
+
+			seenMessageIds.current.add(messageId)
+
+			addMessage({
+				id: messageId,
+				role,
+				content: text || "",
+				toolName,
+				toolDisplayName,
+				toolDisplayOutput,
+				partial,
+				originalType: say,
+				toolData,
+			})
+		},
+		[addMessage],
+	)
+
+	/**
+	 * Process "ask" type messages
+	 */
+	const processAskMessage = useCallback(
+		(messageId: string, ask: ClineAsk, text: string, partial: boolean) => {
+			// DEBUG: Log entry to processAskMessage
+			tuiLogger.debug("ask:process", {
+				ask,
+				partial,
+				nonInteractive,
+				id: messageId,
+			})
+
+			// Handle partial tool asks in nonInteractive mode - stream file content as it arrives
+			// This allows FileWriteTool to show immediately and update as content streams
+			if (partial && ask === "tool" && nonInteractive) {
+				// Parse tool info to extract streaming content
+				let toolName: string | undefined
+				let toolDisplayName: string | undefined
+				let toolDisplayOutput: string | undefined
+				let toolData: ToolData | undefined
+				let parseError = false
+
+				try {
+					const toolInfo = JSON.parse(text) as Record<string, unknown>
+					toolName = toolInfo.tool as string
+					toolDisplayName = toolInfo.tool as string
+					toolDisplayOutput = formatToolOutput(toolInfo)
+					toolData = extractToolData(toolInfo)
+				} catch (err) {
+					// Use raw text if not valid JSON - may happen during early streaming
+					parseError = true
+					tuiLogger.debug("ask:partial-tool:parse-error", {
+						id: messageId,
+						textLen: text.length,
+						textPreview: text.substring(0, 100),
+						error: String(err),
+					})
+				}
+
+				tuiLogger.debug("ask:partial-tool", {
+					id: messageId,
+					textLen: text.length,
+					toolName: toolName || "none",
+					hasToolData: !!toolData,
+					toolDataPath: toolData?.path,
+					toolDataContentLen: toolData?.content?.length || 0,
+					parseError,
+				})
+
+				// Track that we're streaming this tool ask
+				partialToolMessageIdRef.current = messageId
+
+				// Add/update the message with partial content
+				// Use raw JSON text as content so FileWriteTool can parse live content during streaming
+				addMessage({
+					id: messageId,
+					role: "tool",
+					content: text, // Raw JSON text - needed for streaming content parsing
+					toolName,
+					toolDisplayName,
+					toolDisplayOutput,
+					partial: true, // Mark as partial for UI to show loading state
+					originalType: ask,
+					toolData,
+				})
+				return
+			}
+
+			// Skip other partial ask messages - wait for complete
+			if (partial) {
+				return
+			}
+
+			// Skip if already processed (but allow updates to partial tool messages)
+			if (seenMessageIds.current.has(messageId) && partialToolMessageIdRef.current !== messageId) {
+				return
+			}
+
+			// Skip command_output asks (non-blocking)
+			if (ask === "command_output") {
+				seenMessageIds.current.add(messageId)
+				return
+			}
+
+			// Handle resume tasks - don't set pendingAsk
+			if (ask === "resume_task" || ask === "resume_completed_task") {
+				seenMessageIds.current.add(messageId)
+				setLoading(false)
+				useCLIStore.getState().setHasStartedTask(true)
+				useCLIStore.getState().setIsResumingTask(false)
+				return
+			}
+
+			// Track pending command
+			if (ask === "command") {
+				pendingCommandRef.current = text
+			}
+
+			// Handle completion result
+			if (ask === "completion_result") {
+				seenMessageIds.current.add(messageId)
+				// Completion is handled by taskCompleted event
+				// Just add the message for display
+				try {
+					const completionInfo = JSON.parse(text) as Record<string, unknown>
+					const toolData: ToolData = {
+						tool: "attempt_completion",
+						result: completionInfo.result as string | undefined,
+						content: completionInfo.result as string | undefined,
+					}
+
+					addMessage({
+						id: messageId,
+						role: "tool",
+						content: text,
+						toolName: "attempt_completion",
+						toolDisplayName: "Task Complete",
+						toolDisplayOutput: formatToolOutput({ tool: "attempt_completion", ...completionInfo }),
+						originalType: ask,
+						toolData,
+					})
+				} catch {
+					addMessage({
+						id: messageId,
+						role: "tool",
+						content: text || "Task completed",
+						toolName: "attempt_completion",
+						toolDisplayName: "Task Complete",
+						toolDisplayOutput: "✅ Task completed",
+						originalType: ask,
+						toolData: { tool: "attempt_completion", content: text },
+					})
+				}
+				return
+			}
+
+			// For tool/command asks in nonInteractive mode, add as message (auto-approved)
+			if (nonInteractive && ask !== "followup") {
+				seenMessageIds.current.add(messageId)
+
+				if (ask === "tool") {
+					// Clear partial tracking - this is the final message
+					const wasPartial = partialToolMessageIdRef.current === messageId
+					partialToolMessageIdRef.current = null
+
+					let toolName: string | undefined
+					let toolDisplayName: string | undefined
+					let toolDisplayOutput: string | undefined
+					let toolData: ToolData | undefined
+					let todos: TodoItem[] | undefined
+					let previousTodos: TodoItem[] | undefined
+
+					try {
+						const toolInfo = JSON.parse(text) as Record<string, unknown>
+						toolName = toolInfo.tool as string
+						toolDisplayName = toolInfo.tool as string
+						toolDisplayOutput = formatToolOutput(toolInfo)
+						toolData = extractToolData(toolInfo)
+
+						// Handle todo list updates
+						if (toolName === "update_todo_list" || toolName === "updateTodoList") {
+							const parsedTodos = parseTodosFromToolInfo(toolInfo)
+							if (parsedTodos && parsedTodos.length > 0) {
+								todos = parsedTodos
+								previousTodos = [...currentTodos]
+								setTodos(parsedTodos)
+							}
+						}
+					} catch {
+						// Use raw text if not valid JSON
+					}
+
+					addMessage({
+						id: messageId,
+						role: "tool",
+						content: text, // Raw JSON text - needed for tool renderers to parse live content
+						toolName,
+						toolDisplayName,
+						toolDisplayOutput,
+						partial: false, // Final message - not partial
+						originalType: ask,
+						toolData,
+						todos,
+						previousTodos,
+					})
+
+					// If we were streaming, the update already happened via addMessage
+					if (wasPartial) {
+						return
+					}
+				} else if (ask === "command") {
+					// For command asks, add as tool message with command but no output yet
+					// Store the message ID so streaming can update it
+					currentCommandMessageIdRef.current = messageId
+					pendingCommandRef.current = text
+					addMessage({
+						id: messageId,
+						role: "tool",
+						content: "",
+						toolName: "execute_command",
+						toolDisplayName: "bash",
+						originalType: ask,
+						toolData: { tool: "execute_command", command: text },
+					})
+				} else {
+					// Other asks - add as assistant message
+					addMessage({
+						id: messageId,
+						role: "assistant",
+						content: text || "",
+						originalType: ask,
+					})
+				}
+				return
+			}
+
+			// Interactive mode - set pending ask for user input
+			seenMessageIds.current.add(messageId)
+
+			let suggestions: Array<{ answer: string; mode?: string | null }> | undefined
+			let questionText = text
+
+			if (ask === "followup") {
+				try {
+					const data = JSON.parse(text)
+					questionText = data.question || text
+					suggestions = Array.isArray(data.suggest) ? data.suggest : undefined
+				} catch {
+					// Use raw text
+				}
+			} else if (ask === "tool") {
+				try {
+					const toolInfo = JSON.parse(text) as Record<string, unknown>
+					questionText = formatToolAskMessage(toolInfo)
+				} catch {
+					// Use raw text
+				}
+			}
+
+			const pendingAsk: PendingAsk = {
+				id: messageId,
+				type: ask,
+				content: questionText,
+				suggestions,
+			}
+			setPendingAsk(pendingAsk)
+		},
+		[addMessage, setPendingAsk, setLoading, nonInteractive, currentTodos, setTodos],
+	)
+
+	/**
+	 * Handle waitingForInput event from ExtensionClient
+	 * This is emitted when an ask message needs user input
+	 */
+	const handleWaitingForInput = useCallback(
+		(event: WaitingForInputEvent) => {
+			const msg = event.message
+			if (msg.type === "ask" && msg.ask) {
+				processAskMessage(msg.ts.toString(), msg.ask, msg.text || "", false)
+			}
+		},
+		[processAskMessage],
+	)
+
+	// Subscribe to client events
+	useEffect(() => {
+		tuiLogger.debug("useEffect:client", { hasClient: !!client })
+		if (!client) return
+		tuiLogger.debug("useEffect:subscribing", { clientId: "ExtensionClient" })
+
+		// Subscribe to message events
+		const unsubMessage = client.on("message", processClineMessage)
+		const unsubUpdated = client.on("messageUpdated", processClineMessage)
+		const unsubWaiting = client.on("waitingForInput", handleWaitingForInput)
+
+		// Handle streaming terminal output during command execution.
+		// This updates the existing command message with live output.
+		const unsubCommandOutput = client.on("commandExecutionOutput", (event: CommandExecutionOutputEvent) => {
+			// Mark that we've streamed output (to skip the final command_output say message)
+			hasStreamedCommandOutputRef.current = true
+
+			// If we have a command message ID, update that message's output by re-adding with same ID
+			const msgId = currentCommandMessageIdRef.current
+			if (msgId) {
+				// Re-add the message with the same ID to update it (addMessage handles updates)
+				addMessage({
+					id: msgId,
+					role: "tool",
+					content: event.output,
+					toolName: "execute_command",
+					toolDisplayName: "bash",
+					toolDisplayOutput: event.output, // This is what CommandTool displays
+					partial: false, // Non-partial to bypass debounce
+					originalType: "command",
+					toolData: {
+						tool: "execute_command",
+						command: pendingCommandRef.current || undefined,
+						output: event.output,
+					},
+				})
+			} else {
+				// Fallback: create a new message if we don't have a command message ID
+				const streamingMsgId = `streaming-cmd-${event.executionId}`
+				addMessage({
+					id: streamingMsgId,
+					role: "tool",
+					content: event.output,
+					toolName: "execute_command",
+					toolDisplayName: "bash",
+					toolDisplayOutput: event.output,
+					partial: false,
+					originalType: "command_output",
+					toolData: {
+						tool: "execute_command",
+						command: pendingCommandRef.current || undefined,
+						output: event.output,
+					},
+				})
+			}
+		})
+
+		// Update token usage when messages change
+		const unsubStateChange = client.on("stateChange", () => {
+			const messages = client.getMessages()
+			if (messages.length > 1) {
+				const processed = consolidateApiRequests(consolidateCommands(messages.slice(1)))
+				const metrics = consolidateTokenUsage(processed)
+				setTokenUsage(metrics)
+			}
+		})
+
+		return () => {
+			unsubMessage()
+			unsubUpdated()
+			unsubWaiting()
+			unsubCommandOutput()
+			unsubStateChange()
+		}
+	}, [client, processClineMessage, handleWaitingForInput, setTokenUsage])
+
+	return { reset }
+}

apps/cli/src/ui/hooks/useExtensionHost.ts

@@ -1,21 +1,27 @@
-import { useEffect, useRef, useCallback, useMemo } from "react"
+import { useEffect, useRef, useState, useCallback, useMemo } from "react"
 import { useApp } from "ink"
 import { randomUUID } from "crypto"
 import type { ExtensionMessage, WebviewMessage } from "@roo-code/types"
 
-import { ExtensionHostInterface, ExtensionHostOptions } from "@/agent/index.js"
+import { ExtensionClient, ExtensionHostInterface, ExtensionHostOptions } from "@/agent/index.js"
 
 import { useCLIStore } from "../store.js"
 
 export interface UseExtensionHostOptions extends ExtensionHostOptions {
 	initialPrompt?: string
 	exitOnComplete?: boolean
-	onExtensionMessage: (msg: ExtensionMessage) => void
+	/**
+	 * Handle non-message extension state (modes, file search, commands, etc.)
+	 * ClineMessage processing should use useClientEvents instead.
+	 */
+	onExtensionState?: (msg: ExtensionMessage) => void
 	createExtensionHost: (options: ExtensionHostOptions) => ExtensionHostInterface
 }
 
 export interface UseExtensionHostReturn {
 	isReady: boolean
+	/** ExtensionClient for subscribing to message events */
+	client: ExtensionClient | null
 	sendToExtension: ((msg: WebviewMessage) => void) | null
 	runTask: ((prompt: string) => Promise<void>) | null
 	cleanup: () => Promise<void>
@@ -43,19 +49,23 @@ export function useExtensionHost({
 	nonInteractive,
 	ephemeral,
 	exitOnComplete,
-	onExtensionMessage,
+	onExtensionState,
 	createExtensionHost,
 }: UseExtensionHostOptions): UseExtensionHostReturn {
 	const { exit } = useApp()
 	const { addMessage, setComplete, setLoading, setHasStartedTask, setError } = useCLIStore()
 
 	const hostRef = useRef<ExtensionHostInterface | null>(null)
+	// Use state for client so that consumers re-render when it becomes available.
+	// This is critical for useClientEvents which needs the client to subscribe to events.
+	const [client, setClient] = useState<ExtensionClient | null>(null)
 	const isReadyRef = useRef(false)
 
 	const cleanup = useCallback(async () => {
 		if (hostRef.current) {
 			await hostRef.current.dispose()
 			hostRef.current = null
+			setClient(null)
 			isReadyRef.current = false
 		}
 	}, [])
@@ -78,9 +88,15 @@ export function useExtensionHost({
 				})
 
 				hostRef.current = host
+				// Setting client via state triggers re-render so useClientEvents
+				// receives the valid client and can subscribe to events.
+				setClient(host.client)
 				isReadyRef.current = true
 
-				host.on("extensionWebviewMessage", (msg) => onExtensionMessage(msg as ExtensionMessage))
+				// Handle non-message state updates (modes, file search, commands, task history)
+				if (onExtensionState) {
+					host.on("extensionWebviewMessage", (msg) => onExtensionState(msg as ExtensionMessage))
+				}
 
 				host.client.on("taskCompleted", async () => {
 					setComplete(true)
@@ -142,9 +158,9 @@ export function useExtensionHost({
 		return hostRef.current.runTask(prompt)
 	}, [])
 
-	// Memoized return object to prevent unnecessary re-renders in consumers.
+	// Return object includes client state directly so consumers re-render when client changes.
 	return useMemo(
-		() => ({ isReady: isReadyRef.current, sendToExtension, runTask, cleanup }),
-		[sendToExtension, runTask, cleanup],
+		() => ({ isReady: isReadyRef.current, client, sendToExtension, runTask, cleanup }),
+		[client, sendToExtension, runTask, cleanup],
 	)
 }

apps/cli/src/ui/hooks/useExtensionState.ts

@@ -0,0 +1,83 @@
+/**
+ * useExtensionState - Handle non-message extension state updates
+ *
+ * This hook handles extension state that is NOT part of ClineMessage processing:
+ * - Mode changes (current mode, available modes)
+ * - File search results
+ * - Slash commands list
+ * - Task history
+ * - Router models
+ *
+ * ClineMessage processing is handled by useClientEvents, which subscribes to
+ * ExtensionClient events (the unified approach for both TUI and non-TUI modes).
+ */
+
+import { useCallback } from "react"
+import type { ExtensionMessage } from "@roo-code/types"
+
+import type { FileResult, SlashCommandResult, ModeResult } from "../components/autocomplete/index.js"
+import { useCLIStore } from "../store.js"
+
+export interface UseExtensionStateReturn {
+	handleExtensionState: (msg: ExtensionMessage) => void
+}
+
+/**
+ * Hook to handle non-message extension state updates.
+ * This is used alongside useClientEvents which handles ClineMessage events.
+ */
+export function useExtensionState(): UseExtensionStateReturn {
+	const {
+		setFileSearchResults,
+		setAllSlashCommands,
+		setAvailableModes,
+		setCurrentMode,
+		setTaskHistory,
+		setRouterModels,
+	} = useCLIStore()
+
+	/**
+	 * Handle extension messages that contain state updates.
+	 * Only processes non-ClineMessage state.
+	 */
+	const handleExtensionState = useCallback(
+		(msg: ExtensionMessage) => {
+			if (msg.type === "state") {
+				const state = msg.state
+
+				if (!state) {
+					return
+				}
+
+				// Extract and update current mode from state
+				const newMode = state.mode
+
+				if (newMode) {
+					setCurrentMode(newMode)
+				}
+
+				// Extract and update task history from state
+				const newTaskHistory = state.taskHistory
+
+				if (newTaskHistory && Array.isArray(newTaskHistory)) {
+					setTaskHistory(newTaskHistory)
+				}
+
+				// Note: ClineMessages are handled by useClientEvents via ExtensionClient events
+			} else if (msg.type === "fileSearchResults") {
+				setFileSearchResults((msg.results as FileResult[]) || [])
+			} else if (msg.type === "commands") {
+				setAllSlashCommands((msg.commands as SlashCommandResult[]) || [])
+			} else if (msg.type === "modes") {
+				setAvailableModes((msg.modes as ModeResult[]) || [])
+			} else if (msg.type === "routerModels") {
+				if (msg.routerModels) {
+					setRouterModels(msg.routerModels)
+				}
+			}
+		},
+		[setFileSearchResults, setAllSlashCommands, setAvailableModes, setCurrentMode, setTaskHistory, setRouterModels],
+	)
+
+	return { handleExtensionState }
+}

apps/cli/src/ui/hooks/useMessageHandlers.ts

@@ -1,410 +0,0 @@
-import { useCallback, useRef } from "react"
-import type { ExtensionMessage, ClineMessage, ClineAsk, ClineSay, TodoItem } from "@roo-code/types"
-import { consolidateTokenUsage, consolidateApiRequests, consolidateCommands } from "@roo-code/core/cli"
-
-import type { TUIMessage, ToolData } from "../types.js"
-import type { FileResult, SlashCommandResult, ModeResult } from "../components/autocomplete/index.js"
-import { useCLIStore } from "../store.js"
-import { extractToolData, formatToolOutput, formatToolAskMessage, parseTodosFromToolInfo } from "../utils/tools.js"
-
-export interface UseMessageHandlersOptions {
-	nonInteractive: boolean
-}
-
-export interface UseMessageHandlersReturn {
-	handleExtensionMessage: (msg: ExtensionMessage) => void
-	seenMessageIds: React.MutableRefObject<Set<string>>
-	pendingCommandRef: React.MutableRefObject<string | null>
-	firstTextMessageSkipped: React.MutableRefObject<boolean>
-}
-
-/**
- * Hook to handle messages from the extension.
- *
- * Processes three types of messages:
- * 1. "say" messages - Information from the agent (text, tool output, reasoning)
- * 2. "ask" messages - Requests for user input (approvals, followup questions)
- * 3. Extension state updates - Mode changes, task history, file search results
- *
- * Transforms ClineMessage format to TUIMessage format and updates the store.
- */
-export function useMessageHandlers({ nonInteractive }: UseMessageHandlersOptions): UseMessageHandlersReturn {
-	const {
-		addMessage,
-		setPendingAsk,
-		setComplete,
-		setLoading,
-		setHasStartedTask,
-		setFileSearchResults,
-		setAllSlashCommands,
-		setAvailableModes,
-		setCurrentMode,
-		setTokenUsage,
-		setRouterModels,
-		setTaskHistory,
-		currentTodos,
-		setTodos,
-	} = useCLIStore()
-
-	// Track seen message timestamps to filter duplicates and the prompt echo
-	const seenMessageIds = useRef<Set<string>>(new Set())
-	const firstTextMessageSkipped = useRef(false)
-
-	// Track pending command for injecting into command_output toolData
-	const pendingCommandRef = useRef<string | null>(null)
-
-	/**
-	 * Map extension "say" messages to TUI messages
-	 */
-	const handleSayMessage = useCallback(
-		(ts: number, say: ClineSay, text: string, partial: boolean) => {
-			const messageId = ts.toString()
-			const isResuming = useCLIStore.getState().isResumingTask
-
-			if (say === "checkpoint_saved") {
-				return
-			}
-
-			if (say === "api_req_started") {
-				return
-			}
-
-			if (say === "user_feedback") {
-				seenMessageIds.current.add(messageId)
-				return
-			}
-
-			// Skip first text message ONLY for new tasks, not resumed tasks
-			// When resuming, we want to show all historical messages including the first one
-			if (say === "text" && !firstTextMessageSkipped.current && !isResuming) {
-				firstTextMessageSkipped.current = true
-				seenMessageIds.current.add(messageId)
-				return
-			}
-
-			if (seenMessageIds.current.has(messageId) && !partial) {
-				return
-			}
-
-			let role: TUIMessage["role"] = "assistant"
-			let toolName: string | undefined
-			let toolDisplayName: string | undefined
-			let toolDisplayOutput: string | undefined
-			let toolData: ToolData | undefined
-
-			if (say === "command_output") {
-				role = "tool"
-				toolName = "execute_command"
-				toolDisplayName = "bash"
-				toolDisplayOutput = text
-				const trackedCommand = pendingCommandRef.current
-				toolData = { tool: "execute_command", command: trackedCommand || undefined, output: text }
-				pendingCommandRef.current = null
-			} else if (say === "reasoning") {
-				role = "thinking"
-			}
-
-			seenMessageIds.current.add(messageId)
-
-			addMessage({
-				id: messageId,
-				role,
-				content: text || "",
-				toolName,
-				toolDisplayName,
-				toolDisplayOutput,
-				partial,
-				originalType: say,
-				toolData,
-			})
-		},
-		[addMessage],
-	)
-
-	/**
-	 * Handle extension "ask" messages
-	 */
-	const handleAskMessage = useCallback(
-		(ts: number, ask: ClineAsk, text: string, partial: boolean) => {
-			const messageId = ts.toString()
-
-			if (partial) {
-				return
-			}
-
-			if (seenMessageIds.current.has(messageId)) {
-				return
-			}
-
-			if (ask === "command_output") {
-				seenMessageIds.current.add(messageId)
-				return
-			}
-
-			// Handle resume_task and resume_completed_task - stop loading and show text input
-			// Do not set pendingAsk - just stop loading so user sees normal input to type new message
-			if (ask === "resume_task" || ask === "resume_completed_task") {
-				seenMessageIds.current.add(messageId)
-				setLoading(false)
-				// Mark that a task has been started so subsequent messages continue the task
-				// (instead of starting a brand new task via runTask)
-				setHasStartedTask(true)
-				// Clear the resuming flag since we're now ready for interaction
-				// Historical messages should already be displayed from state processing
-				useCLIStore.getState().setIsResumingTask(false)
-				// Do not set pendingAsk - let the normal text input appear
-				return
-			}
-
-			if (ask === "completion_result") {
-				seenMessageIds.current.add(messageId)
-				setComplete(true)
-				setLoading(false)
-
-				// Parse the completion result and add a message for CompletionTool to render
-				try {
-					const completionInfo = JSON.parse(text) as Record<string, unknown>
-					const toolData: ToolData = {
-						tool: "attempt_completion",
-						result: completionInfo.result as string | undefined,
-						content: completionInfo.result as string | undefined,
-					}
-
-					addMessage({
-						id: messageId,
-						role: "tool",
-						content: text,
-						toolName: "attempt_completion",
-						toolDisplayName: "Task Complete",
-						toolDisplayOutput: formatToolOutput({ tool: "attempt_completion", ...completionInfo }),
-						originalType: ask,
-						toolData,
-					})
-				} catch {
-					// If parsing fails, still add a basic completion message
-					addMessage({
-						id: messageId,
-						role: "tool",
-						content: text || "Task completed",
-						toolName: "attempt_completion",
-						toolDisplayName: "Task Complete",
-						toolDisplayOutput: "✅ Task completed",
-						originalType: ask,
-						toolData: {
-							tool: "attempt_completion",
-							content: text,
-						},
-					})
-				}
-				return
-			}
-
-			// Track pending command BEFORE nonInteractive handling
-			// This ensures we capture the command text for later injection into command_output toolData
-			if (ask === "command") {
-				pendingCommandRef.current = text
-			}
-
-			if (nonInteractive && ask !== "followup") {
-				seenMessageIds.current.add(messageId)
-
-				if (ask === "tool") {
-					let toolName: string | undefined
-					let toolDisplayName: string | undefined
-					let toolDisplayOutput: string | undefined
-					let formattedContent = text || ""
-					let toolData: ToolData | undefined
-					let todos: TodoItem[] | undefined
-					let previousTodos: TodoItem[] | undefined
-
-					try {
-						const toolInfo = JSON.parse(text) as Record<string, unknown>
-						toolName = toolInfo.tool as string
-						toolDisplayName = toolInfo.tool as string
-						toolDisplayOutput = formatToolOutput(toolInfo)
-						formattedContent = formatToolAskMessage(toolInfo)
-						// Extract structured toolData for rich rendering
-						toolData = extractToolData(toolInfo)
-
-						// Special handling for update_todo_list tool - extract todos
-						if (toolName === "update_todo_list" || toolName === "updateTodoList") {
-							const parsedTodos = parseTodosFromToolInfo(toolInfo)
-							if (parsedTodos && parsedTodos.length > 0) {
-								todos = parsedTodos
-								// Capture previous todos before updating global state
-								previousTodos = [...currentTodos]
-								setTodos(parsedTodos)
-							}
-						}
-					} catch {
-						// Use raw text if not valid JSON
-					}
-
-					addMessage({
-						id: messageId,
-						role: "tool",
-						content: formattedContent,
-						toolName,
-						toolDisplayName,
-						toolDisplayOutput,
-						originalType: ask,
-						toolData,
-						todos,
-						previousTodos,
-					})
-				} else {
-					addMessage({
-						id: messageId,
-						role: "assistant",
-						content: text || "",
-						originalType: ask,
-					})
-				}
-				return
-			}
-
-			let suggestions: Array<{ answer: string; mode?: string | null }> | undefined
-			let questionText = text
-
-			if (ask === "followup") {
-				try {
-					const data = JSON.parse(text)
-					questionText = data.question || text
-					suggestions = Array.isArray(data.suggest) ? data.suggest : undefined
-				} catch {
-					// Use raw text
-				}
-			} else if (ask === "tool") {
-				try {
-					const toolInfo = JSON.parse(text) as Record<string, unknown>
-					questionText = formatToolAskMessage(toolInfo)
-				} catch {
-					// Use raw text if not valid JSON
-				}
-			}
-			// Note: ask === "command" is handled above before the nonInteractive block
-
-			seenMessageIds.current.add(messageId)
-
-			setPendingAsk({
-				id: messageId,
-				type: ask,
-				content: questionText,
-				suggestions,
-			})
-		},
-		[addMessage, setPendingAsk, setComplete, setLoading, setHasStartedTask, nonInteractive, currentTodos, setTodos],
-	)
-
-	/**
-	 * Handle all extension messages
-	 */
-	const handleExtensionMessage = useCallback(
-		(msg: ExtensionMessage) => {
-			if (msg.type === "state") {
-				const state = msg.state
-
-				if (!state) {
-					return
-				}
-
-				// Extract and update current mode from state
-				const newMode = state.mode
-
-				if (newMode) {
-					setCurrentMode(newMode)
-				}
-
-				// Extract and update task history from state
-				const newTaskHistory = state.taskHistory
-
-				if (newTaskHistory && Array.isArray(newTaskHistory)) {
-					setTaskHistory(newTaskHistory)
-				}
-
-				const clineMessages = state.clineMessages
-
-				if (clineMessages) {
-					for (const clineMsg of clineMessages) {
-						const ts = clineMsg.ts
-						const type = clineMsg.type
-						const say = clineMsg.say
-						const ask = clineMsg.ask
-						const text = clineMsg.text || ""
-						const partial = clineMsg.partial || false
-
-						if (type === "say" && say) {
-							handleSayMessage(ts, say, text, partial)
-						} else if (type === "ask" && ask) {
-							handleAskMessage(ts, ask, text, partial)
-						}
-					}
-
-					// Compute token usage metrics from clineMessages
-					// Skip first message (task prompt) as per webview UI pattern
-					if (clineMessages.length > 1) {
-						const processed = consolidateApiRequests(
-							consolidateCommands(clineMessages.slice(1) as ClineMessage[]),
-						)
-
-						const metrics = consolidateTokenUsage(processed)
-						setTokenUsage(metrics)
-					}
-				}
-
-				// After processing state, clear the resuming flag if it was set
-				// This ensures the flag is cleared even if no resume_task ask message is received
-				if (useCLIStore.getState().isResumingTask) {
-					useCLIStore.getState().setIsResumingTask(false)
-				}
-			} else if (msg.type === "messageUpdated") {
-				const clineMessage = msg.clineMessage
-
-				if (!clineMessage) {
-					return
-				}
-
-				const ts = clineMessage.ts
-				const type = clineMessage.type
-				const say = clineMessage.say
-				const ask = clineMessage.ask
-				const text = clineMessage.text || ""
-				const partial = clineMessage.partial || false
-
-				if (type === "say" && say) {
-					handleSayMessage(ts, say, text, partial)
-				} else if (type === "ask" && ask) {
-					handleAskMessage(ts, ask, text, partial)
-				}
-			} else if (msg.type === "fileSearchResults") {
-				setFileSearchResults((msg.results as FileResult[]) || [])
-			} else if (msg.type === "commands") {
-				setAllSlashCommands((msg.commands as SlashCommandResult[]) || [])
-			} else if (msg.type === "modes") {
-				setAvailableModes((msg.modes as ModeResult[]) || [])
-			} else if (msg.type === "routerModels") {
-				if (msg.routerModels) {
-					setRouterModels(msg.routerModels)
-				}
-			}
-		},
-		[
-			handleSayMessage,
-			handleAskMessage,
-			setFileSearchResults,
-			setAllSlashCommands,
-			setAvailableModes,
-			setCurrentMode,
-			setTokenUsage,
-			setRouterModels,
-			setTaskHistory,
-		],
-	)
-
-	return {
-		handleExtensionMessage,
-		seenMessageIds,
-		pendingCommandRef,
-		firstTextMessageSkipped,
-	}
-}

apps/cli/src/ui/store.ts

@@ -1,10 +1,13 @@
 import { create } from "zustand"
 
 import type { TokenUsage, ProviderSettings, TodoItem } from "@roo-code/types"
+import { DebugLogger } from "@roo-code/core/cli"
 
 import type { TUIMessage, PendingAsk, TaskHistoryItem } from "./types.js"
 import type { FileResult, SlashCommandResult, ModeResult } from "./components/autocomplete/index.js"
 
+const storeLogger = new DebugLogger("STORE")
+
 /**
  * Shallow array equality check - compares array length and element references.
  * Used to prevent unnecessary state updates when array content hasn't changed.
@@ -162,10 +165,24 @@ export const useCLIStore = create<CLIState & CLIActions>((set, get) => ({
 
 		// For NEW messages (not updates) - always apply immediately
 		if (existingIndex === -1) {
+			storeLogger.debug("addMessage:new", {
+				id: msg.id,
+				role: msg.role,
+				toolName: msg.toolName || "none",
+				partial: msg.partial,
+				hasToolData: !!msg.toolData,
+				msgCount: state.messages.length + 1,
+			})
 			set({ messages: [...state.messages, msg] })
 			return
 		}
 
+		storeLogger.debug("addMessage:update", {
+			id: msg.id,
+			partial: msg.partial,
+			existingIndex,
+		})
+
 		// For UPDATES to existing messages:
 		// If partial (streaming) and message exists, debounce the update
 		if (msg.partial) {

apps/cli/tsup.config.ts

@@ -11,8 +11,18 @@ export default defineConfig({
 	banner: {
 		js: "#!/usr/bin/env node",
 	},
-	// Bundle workspace packages that export TypeScript
-	noExternal: ["@roo-code/core", "@roo-code/core/cli", "@roo-code/types", "@roo-code/vscode-shim"],
+	// Bundle workspace packages and ESM-only npm dependencies to create a self-contained CLI
+	noExternal: [
+		// Workspace packages
+		"@roo-code/core",
+		"@roo-code/core/cli",
+		"@roo-code/types",
+		"@roo-code/vscode-shim",
+		// ESM-only npm dependencies that need to be bundled
+		"@agentclientprotocol/sdk",
+		"p-wait-for",
+		"zod",
+	],
 	external: [
 		// Keep native modules external
 		"@anthropic-ai/sdk",

apps/web-roo-code/src/app/cloud/page.tsx

@@ -98,8 +98,7 @@ const features: Feature[] = [
 	{
 		icon: Brain,
 		title: "Model Agnostic",
-		description:
-			"Bring your own keys or use the Roo Code Router with access to all top models with no markup.",
+		description: "Bring your own keys or use the Roo Code Router with access to all top models with no markup.",
 	},
 	{
 		icon: Github,
@@ -153,7 +152,7 @@ export default function CloudPage() {
 							Your AI Team <span className="text-violet-500">in the Cloud</span>
 						</h1>
 						<p className="text-xl text-muted-foreground mb-8 max-w-2xl mx-auto">
-						Create your agent team in the Cloud, give them access to GitHub, and start delegating tasks
+							Create your agent team in the Cloud, give them access to GitHub, and start delegating tasks
 							from Web and Slack.
 						</p>
 						<div className="flex flex-col sm:flex-row gap-4 justify-center">

apps/web-roo-code/src/app/pricing/page.tsx

@@ -234,8 +234,8 @@ export default function PricingPage() {
 							<div className="text-sm text-muted-foreground">
 								<p className="">
 									On any plan, you can use your own LLM provider API key or use the built-in Roo Code
-									Router – curated models to work with Roo with no markup, including the
-									latest Gemini, GPT and Claude. Paid with credits.
+									Router – curated models to work with Roo with no markup, including the latest
+									Gemini, GPT and Claude. Paid with credits.
 									<Link href="/provider" className="underline hover:no-underline ml-1">
 										See per model pricing.
 									</Link>

pnpm-lock.yaml

@@ -82,6 +82,9 @@ importers:
 
   apps/cli:
     dependencies:
+      '@agentclientprotocol/sdk':
+        specifier: ^0.12.0
+        version: 0.12.0(zod@4.3.5)
       '@inkjs/ui':
         specifier: ^2.0.0
         version: 2.0.0(ink@6.6.0(@types/react@18.3.23)(react@19.2.3))
@@ -118,6 +121,9 @@ importers:
       superjson:
         specifier: ^2.2.6
         version: 2.2.6
+      zod:
+        specifier: ^4.3.5
+        version: 4.3.5
       zustand:
         specifier: ^5.0.0
         version: 5.0.9(@types/react@18.3.23)(react@19.2.3)
@@ -1356,6 +1362,11 @@ packages:
   '@adobe/css-tools@4.4.2':
     resolution: {integrity: sha512-baYZExFpsdkBNuvGKTKWCwKH57HRZLVtycZS05WTQNVOiXVSeAki3nU35zlRbToeMW8aHlJfyS+1C4BOv27q0A==}
 
+  '@agentclientprotocol/sdk@0.12.0':
+    resolution: {integrity: sha512-V8uH/KK1t7utqyJmTA7y7DzKu6+jKFIXM+ZVouz8E55j8Ej2RV42rEvPKn3/PpBJlliI5crcGk1qQhZ7VwaepA==}
+    peerDependencies:
+      zod: ^3.25.0 || ^4.0.0
+
   '@alcalzone/ansi-tokenize@0.2.3':
     resolution: {integrity: sha512-jsElTJ0sQ4wHRz+C45tfect76BwbTbgkgKByOzpCN9xG61N5V6u/glvg1CsNJhq2xJIFpKHSwG3D2wPPuEYOrQ==}
     engines: {node: '>=18'}
@@ -10659,6 +10670,9 @@ packages:
   zod@3.25.76:
     resolution: {integrity: sha512-gzUt/qt81nXsFGKIFcC3YnfEAx5NkunCfnDlvuBSSFS02bcXu4Lmea0AFIUwbLWxWPx3d9p8S5QoaujKcNQxcQ==}
 
+  zod@4.3.5:
+    resolution: {integrity: sha512-k7Nwx6vuWx1IJ9Bjuf4Zt1PEllcwe7cls3VNzm4CQ1/hgtFUK2bRNG3rvnpPUhFjmqJKAKtjV576KnUkHocg/g==}
+
   zustand@5.0.9:
     resolution: {integrity: sha512-ALBtUj0AfjJt3uNRQoL1tL2tMvj6Gp/6e39dnfT6uzpelGru8v1tPOGBzayOWbPJvujM8JojDk3E1LxeFisBNg==}
     engines: {node: '>=12.20.0'}
@@ -10684,6 +10698,10 @@ snapshots:
 
   '@adobe/css-tools@4.4.2': {}
 
+  '@agentclientprotocol/sdk@0.12.0(zod@4.3.5)':
+    dependencies:
+      zod: 4.3.5
+
   '@alcalzone/ansi-tokenize@0.2.3':
     dependencies:
       ansi-styles: 6.2.3
@@ -21715,6 +21733,8 @@ snapshots:
 
   zod@3.25.76: {}
 
+  zod@4.3.5: {}
+
   zustand@5.0.9(@types/react@18.3.23)(react@19.2.3):
     optionalDependencies:
       '@types/react': 18.3.23

src/core/tools/WriteToFileTool.ts

@@ -26,6 +26,12 @@ interface WriteToFileParams {
 export class WriteToFileTool extends BaseTool<"write_to_file"> {
 	readonly name = "write_to_file" as const
 
+	/**
+	 * Track whether we've sent the initial "tool starting" notification.
+	 * This allows us to send an immediate notification before path stabilizes.
+	 */
+	private hasNotifiedToolStart = false
+
 	parseLegacy(params: Partial<Record<string, string>>): WriteToFileParams {
 		return {
 			path: params.path || "",
@@ -124,6 +130,10 @@ export class WriteToFileTool extends BaseTool<"write_to_file"> {
 					task.diffViewProvider.originalContent = ""
 				}
 
+				// Send partial message immediately to indicate tool is starting (before file write)
+				const partialMessage = JSON.stringify(sharedMessageProps)
+				await task.ask("tool", partialMessage, true).catch(() => {})
+
 				let unified = fileExists
 					? formatResponse.createPrettyPatch(relPath, task.diffViewProvider.originalContent, newContent)
 					: convertNewFileToUnifiedDiff(newContent, relPath)
@@ -200,11 +210,33 @@ export class WriteToFileTool extends BaseTool<"write_to_file"> {
 		}
 	}
 
+	/**
+	 * Reset partial state tracking, including the tool start notification flag.
+	 */
+	override resetPartialState(): void {
+		super.resetPartialState()
+		this.hasNotifiedToolStart = false
+	}
+
 	override async handlePartial(task: Task, block: ToolUse<"write_to_file">): Promise<void> {
 		const relPath: string | undefined = block.params.path
 		let newContent: string | undefined = block.params.content
 
-		// Wait for path to stabilize before showing UI (prevents truncated paths)
+		// Send an immediate "tool starting" notification on first partial call
+		// This ensures CLI sees the tool start immediately, before path stabilizes
+		if (!this.hasNotifiedToolStart && relPath) {
+			this.hasNotifiedToolStart = true
+			const startMessage: ClineSayTool = {
+				tool: "newFileCreated", // Will be updated when we know if file exists
+				path: relPath,
+				content: "",
+				isOutsideWorkspace: false,
+				isProtected: false,
+			}
+			await task.ask("tool", JSON.stringify(startMessage), true).catch(() => {})
+		}
+
+		// Wait for path to stabilize before showing full UI (prevents truncated paths)
 		if (!this.hasPathStabilized(relPath) || newContent === undefined) {
 			return
 		}
@@ -216,10 +248,6 @@ export class WriteToFileTool extends BaseTool<"write_to_file"> {
 			EXPERIMENT_IDS.PREVENT_FOCUS_DISRUPTION,
 		)
 
-		if (isPreventFocusDisruptionEnabled) {
-			return
-		}
-
 		// relPath is guaranteed non-null after hasPathStabilized
 		let fileExists: boolean
 		const absolutePath = path.resolve(task.cwd, relPath!)
@@ -248,9 +276,15 @@ export class WriteToFileTool extends BaseTool<"write_to_file"> {
 			isProtected: isWriteProtected,
 		}
 
+		// Always send partial messages to keep CLI informed during streaming
 		const partialMessage = JSON.stringify(sharedMessageProps)
 		await task.ask("tool", partialMessage, block.partial).catch(() => {})
 
+		// Skip diff view operations when experiment is enabled (prevents focus disruption in VSCode)
+		if (isPreventFocusDisruptionEnabled) {
+			return
+		}
+
 		if (newContent) {
 			if (!task.diffViewProvider.isEditing) {
 				await task.diffViewProvider.open(relPath!)

src/core/tools/__tests__/writeToFileTool.spec.ts

@@ -400,14 +400,14 @@ describe("writeToFileTool", () => {
 		})
 
 		it("streams content updates during partial execution after path stabilizes", async () => {
-			// First call - path not yet stabilized, early return (no file operations)
+			// First call - sends early "tool starting" notification, but no file operations yet
 			await executeWriteFileTool({}, { isPartial: true })
-			expect(mockCline.ask).not.toHaveBeenCalled()
+			expect(mockCline.ask).toHaveBeenCalledTimes(1) // Early notification sent
 			expect(mockCline.diffViewProvider.open).not.toHaveBeenCalled()
 
 			// Second call with same path - path is now stabilized, file operations proceed
 			await executeWriteFileTool({}, { isPartial: true })
-			expect(mockCline.ask).toHaveBeenCalled()
+			expect(mockCline.ask).toHaveBeenCalledTimes(2) // Additional call after path stabilizes
 			expect(mockCline.diffViewProvider.open).toHaveBeenCalledWith(testFilePath)
 			expect(mockCline.diffViewProvider.update).toHaveBeenCalledWith(testContent, false)
 		})