claude: Phase 10 — workbench client tools via in-process MCP server

.agents/skills/launch/SKILL.md

@@ -128,6 +128,16 @@ npx @playwright/cli snapshot
 - The `--agents` flag launches the Agents workbench instead of the standard VS Code workbench.
 - Set `VSCODE_SKIP_PRELAUNCH=1` to skip the compile step if you've already built.
 
+> **🚨 Driving the Agents window UI is not the same as driving regular Code OSS.**
+> The session sidebar, agent picker, and chat input have their own quirks that
+> trip up the obvious `click + type` patterns. Before you spend time fighting
+> the UI, read **[references/agents-window-guide.md](./references/agents-window-guide.md)**
+> — it covers the agent picker (Copilot CLI vs Claude), `[Local]` vs non-`[Local]`
+> sessions, the reliable keyboard-navigation trick for the agent dropdown, how
+> to send a follow-up turn in the same session, how to restore an existing
+> chat, and the kill / relaunch / verify-out-is-fresh dance you need after a
+> source code change.
+
 ## Launching VS Code Extensions for Debugging
 
 To debug a VS Code extension via npx @playwright/cli, launch VS Code Insiders with `--extensionDevelopmentPath` and `--remote-debugging-port`. Use `--user-data-dir` to avoid conflicting with an already-running instance.

eslint.config.js

@@ -1503,6 +1503,7 @@ export default defineConfig(
 					'when': 'hasNode',
 					'allow': [
 						'@github/copilot-sdk',
+						'zod',
 						'@microsoft/dev-tunnels-contracts',
 						'@microsoft/dev-tunnels-management',
 						'@parcel/watcher',
@@ -1646,7 +1647,9 @@ export default defineConfig(
 						'@vscode/tree-sitter-wasm', // used by agentHost for command auto-approval
 						'@vscode/copilot-api', // used by agentHost for Copilot API requests
 						'@anthropic-ai/sdk', // used by agentHost for Anthropic API requests
-						'@anthropic-ai/claude-agent-sdk' // used by agentHost for Claude Agent SDK session enumeration / queries
+						'@anthropic-ai/claude-agent-sdk', // used by agentHost for Claude Agent SDK session enumeration / queries
+						'@modelcontextprotocol/sdk/**/*', // used by agentHost for Claude client-tool MCP result types (Phase 10)
+						'zod' // used by agentHost for Claude client-tool MCP input schemas
 					]
 				},
 				{

package.json

@@ -147,7 +147,8 @@
     "vscode-textmate": "^9.3.2",
     "ws": "^8.19.0",
     "yauzl": "^3.0.0",
-    "yazl": "^2.4.3"
+    "yazl": "^2.4.3",
+    "zod": "^3.25.76"
   },
   "devDependencies": {
     "@anthropic-ai/claude-agent-sdk": "0.2.128",

remote/package.json

@@ -46,7 +46,8 @@
     "vscode-textmate": "^9.3.2",
     "ws": "^8.19.0",
     "yauzl": "^3.0.0",
-    "yazl": "^2.4.3"
+    "yazl": "^2.4.3",
+    "zod": "^3.25.76"
   },
   "overrides": {
     "node-gyp-build": "4.8.1",

src/vs/platform/agentHost/common/pendingRequestRegistry.ts

@@ -4,6 +4,7 @@
  *--------------------------------------------------------------------------------------------*/
 
 import { DeferredPromise } from '../../../base/common/async.js';
+import { CancellationError } from '../../../base/common/errors.js';
 
 /**
  * Registry of parked deferred promises keyed by string id. Used to
@@ -27,6 +28,28 @@ export class PendingRequestRegistry<T> {
 		return deferred.p;
 	}
 
+	/**
+	 * Park a deferred under `key` and return its promise. Use when there
+	 * is no synchronous responder to guard against — the request that
+	 * eventually feeds {@link respond} originates from a different code
+	 * path (e.g. an MCP handler invoked by the SDK whose completion
+	 * arrives via a workbench round-trip).
+	 *
+	 * If `key` is already registered (duplicate `tool_use_id` from the
+	 * SDK, retry, or logic bug), the previous deferred is rejected with
+	 * a {@link CancellationError} so its awaiter unwinds instead of
+	 * leaking forever.
+	 */
+	register(key: string): Promise<T> {
+		const existing = this._entries.get(key);
+		if (existing && !existing.isSettled) {
+			existing.error(new CancellationError());
+		}
+		const deferred = new DeferredPromise<T>();
+		this._entries.set(key, deferred);
+		return deferred.p;
+	}
+
 	respond(key: string, value: T): boolean {
 		const deferred = this._entries.get(key);
 		if (!deferred) {
@@ -37,6 +60,14 @@ export class PendingRequestRegistry<T> {
 		return true;
 	}
 
+	/**
+	 * Resolve every parked deferred with `denyValue` and clear the registry.
+	 *
+	 * Designed for the permission-deny path: a "deny" answer is itself a
+	 * successful round-trip result, so awaiting consumers receive `denyValue`
+	 * rather than an error. Use {@link rejectAll} when callers must observe
+	 * a thrown error instead (cancellation, dispose).
+	 */
 	denyAll(denyValue: T): void {
 		for (const [, deferred] of this._entries) {
 			if (!deferred.isSettled) {
@@ -45,4 +76,23 @@ export class PendingRequestRegistry<T> {
 		}
 		this._entries.clear();
 	}
+
+	/**
+	 * Reject every parked deferred with `error` and clear the registry.
+	 *
+	 * Use this when in-flight requests must be cancelled rather than
+	 * answered (e.g. session dispose, `Query` rebind on tool-set change).
+	 * Compare with {@link denyAll}, which *resolves* every deferred with a
+	 * supplied value — that is right for the permission-deny path where a
+	 * "deny" is itself a successful answer, but wrong for cancellation
+	 * where the awaited consumer must observe an error to unwind.
+	 */
+	rejectAll(error: Error): void {
+		for (const [, deferred] of this._entries) {
+			if (!deferred.isSettled) {
+				deferred.error(error);
+			}
+		}
+		this._entries.clear();
+	}
 }