@shanepadgett/tau-agent 0.43.0 → 0.44.0

package/extensions/script-runner/index.ts

@@ -169,7 +169,6 @@ export default function scriptRunnerExtension(pi: ExtensionAPI): void {
 	const langPhrase = formatLangList(detected);
 
 	const scripts = new Map<string, StoredScript>();
-	let tempDir: string | undefined;
 
 	function remember(scriptId: string, script: StoredScript): void {
 		scripts.set(scriptId, script);
@@ -190,12 +189,6 @@ export default function scriptRunnerExtension(pi: ExtensionAPI): void {
 		throw new Error("Deno is not available on this machine.");
 	}
 
-	async function ensureTempDir(): Promise<string> {
-		if (tempDir) return tempDir;
-		tempDir = await mkdtemp(join(tmpdir(), "tau-script-runner-"));
-		return tempDir;
-	}
-
 	async function runScript(
 		language: Language,
 		command: string,
@@ -203,21 +196,27 @@ export default function scriptRunnerExtension(pi: ExtensionAPI): void {
 		cwd: string,
 		signal: AbortSignal | undefined,
 	): Promise<ExecResult> {
-		const dir = await ensureTempDir();
-		const file = join(dir, language === "python3" ? "_run.py" : "_run.ts");
-		await writeFile(file, source, "utf8");
-		const args =
-			language === "python3"
-				? [file]
-				: language === "node"
-					? ["--experimental-strip-types", file]
-					: ["run", "-A", file];
-		const result = await pi.exec(command, args, { cwd, signal, timeout: TIMEOUT_MS });
-		return {
-			...result,
-			stdout: scrubPath(result.stdout, file, dir),
-			stderr: scrubPath(result.stderr, file, dir),
-		};
+		const dir = await mkdtemp(join(tmpdir(), "tau-script-runner-"));
+		try {
+			const file = join(dir, language === "python3" ? "_run.py" : "_run.ts");
+			await writeFile(file, source, "utf8");
+			const args =
+				language === "python3"
+					? [file]
+					: language === "node"
+						? ["--experimental-strip-types", file]
+						: ["run", "-A", file];
+			const result = await pi.exec(command, args, { cwd, signal, timeout: TIMEOUT_MS });
+			return {
+				...result,
+				stdout: scrubPath(result.stdout, file, dir),
+				stderr: scrubPath(result.stderr, file, dir),
+			};
+		} finally {
+			await rm(dir, { recursive: true, force: true }).catch(() => {
+				// best-effort cleanup
+			});
+		}
 	}
 
 	const paramsSchema = Type.Object(
@@ -315,12 +314,5 @@ export default function scriptRunnerExtension(pi: ExtensionAPI): void {
 
 	pi.on("session_shutdown", () => {
 		scripts.clear();
-		const dir = tempDir;
-		tempDir = undefined;
-		if (dir) {
-			void rm(dir, { recursive: true, force: true }).catch(() => {
-				// best-effort cleanup
-			});
-		}
 	});
 }

package/extensions/tool-approval/README.md

@@ -2,11 +2,11 @@
 
 Reviews agent `bash` and `script_runner` requests before they run.
 
-Common read-only bash commands skip review and run immediately. Other bash and every `script_runner` request go to a separate reviewer. Tau uses the reviewer model for the current provider, then the current chat model if that reviewer is unavailable or fails. If a reviewer model is unavailable or fails, Tau notifies and tries the next one. The reviewer returns a validated decision and one concise paragraph that explains the request.
+Common read-only bash commands skip review and run immediately. Other bash and every `script_runner` request go to a separate reviewer. When the agent requests several tools at once, Tau reviews up to three requests concurrently, with one model review focused on each request. Tau uses the reviewer model for the current provider, then the current chat model if that reviewer is unavailable or fails. If a reviewer model is unavailable or fails, Tau notifies and tries the next one. The reviewer returns a validated decision and one concise paragraph that explains the request.
 
 With `autoApprove` enabled, reviewer-approved requests run without another confirmation. Tau shows a user-only marker with the reviewer model after those auto-approvals. Common read-only bash that skips review does not get a marker. Routine local development work should be approved, including requests that modify project files or run scripts. The reviewer asks for human approval only when it finds a concrete destructive, system, production, privileged, or security-sensitive effect.
 
-When approval is required, Tau shows one paragraph that explains the effect and risk without repeating the request. If the reviewer fails or returns a malformed decision, Tau asks for direct human approval instead of running it automatically. Tau also sends an attention notification when the approval window opens.
+When approval is required, Tau shows one paragraph that explains the effect and risk without repeating the request. If several requests need approval, their confirmation windows open one at a time. If the reviewer fails or returns a malformed decision, Tau asks for direct human approval instead of running it automatically. Tau also sends an attention notification when the approval window opens.
 
 In the terminal approval panel, move between Approve and Reject, press `n` to add a note to the highlighted choice, then press Enter to choose. Enter saves an edited note before choosing; Escape cancels note editing or blocks the request from the choice list. A rejection note tells the agent why the request was blocked. An approval note reaches the agent with the tool result; it does not change the request being approved. To ask for a different request, reject it with a note. Long notes are truncated. RPC clients use the standard confirmation dialog without notes.
 

package/extensions/tool-approval/index.ts

@@ -17,6 +17,7 @@ import toolApprovalSettings from "./settings.ts";
 
 const STATUS_KEY = "tool-approval";
 const AUTO_APPROVED_TYPE = "tau.tool-approval.auto-approved";
+const MAX_CONCURRENT_REVIEWS = 3;
 
 const REVIEW_SCHEMA = Type.Object(
 	{
@@ -75,6 +76,13 @@ interface ToolApprovalRequest {
 	input: Record<string, unknown>;
 }
 
+type ToolReviewResult = { review: ToolReview; provider: string; model: string };
+
+interface CachedReview {
+	requestJson: string;
+	outcome: PromiseSettledResult<ToolReviewResult>;
+}
+
 interface AutoApprovedMarker {
 	toolName: ApprovalToolName;
 	provider: string;
@@ -83,6 +91,7 @@ interface AutoApprovedMarker {
 
 export default function toolApprovalExtension(pi: ExtensionAPI): void {
 	let settings = toolApprovalSettings.defaults;
+	let batchReviews: Map<string, CachedReview> | undefined;
 	const pendingNotes = new Map<string, string>();
 
 	pi.registerEntryRenderer<AutoApprovedMarker>(AUTO_APPROVED_TYPE, (entry, _options, theme) => {
@@ -136,6 +145,7 @@ export default function toolApprovalExtension(pi: ExtensionAPI): void {
 	}
 
 	pi.on("session_start", async (_event, ctx) => {
+		batchReviews = undefined;
 		pendingNotes.clear();
 		await refreshSettings(ctx);
 	});
@@ -152,21 +162,32 @@ export default function toolApprovalExtension(pi: ExtensionAPI): void {
 		}
 		if (!settings.enabled) return undefined;
 
-		let command: string | undefined;
 		if (request.toolName === "bash") {
-			const value = request.input.command;
-			if (typeof value !== "string") return block("bash command was malformed");
-			if (!value.trim()) {
+			const command = request.input.command;
+			if (typeof command !== "string") return block("bash command was malformed");
+			if (!command.trim()) {
 				ctx.ui.notify("Bash command blocked: command is empty", "warning");
 				return block("bash command is empty");
 			}
-			command = value;
 			if (isAllowlistedBash(command)) return undefined;
 		}
 
 		ctx.ui.setStatus(STATUS_KEY, `reviewing ${toolLabel(request.toolName)}`);
 		try {
-			const { review, provider, model } = await reviewToolRequest(ctx, request);
+			batchReviews ??= await reviewAssistantRequests(ctx, event, request);
+			if (ctx.signal?.aborted) return block("Tool review cancelled");
+			const cached = batchReviews.get(event.toolCallId);
+			batchReviews.delete(event.toolCallId);
+			let result: ToolReviewResult;
+			if (cached && cached.requestJson === JSON.stringify(request)) {
+				if (cached.outcome.status === "rejected") throw cached.outcome.reason;
+				result = cached.outcome.value;
+			} else {
+				// A sibling's validated input may differ from its arguments in the assistant message.
+				result = await reviewToolRequest(ctx, request);
+			}
+			if (ctx.signal?.aborted) return block("Tool review cancelled");
+			const { review, provider, model } = result;
 			if (review.decision === "requires_user_approval") {
 				return requestToolApproval(
 					ctx,
@@ -192,6 +213,7 @@ export default function toolApprovalExtension(pi: ExtensionAPI): void {
 				formatApproval(review.summary, "Automatic approval is disabled."),
 			);
 		} catch (error) {
+			if (ctx.signal?.aborted) return block("Tool review cancelled");
 			const message = singleLine(errorText(error));
 			ctx.ui.notify(`Tool review failed; manual approval required: ${truncAt(message, 600)}`, "warning");
 			return requestToolApproval(
@@ -206,6 +228,10 @@ export default function toolApprovalExtension(pi: ExtensionAPI): void {
 		}
 	});
 
+	pi.on("turn_end", () => {
+		batchReviews = undefined;
+	});
+
 	pi.on("tool_result", (event) => {
 		const note = pendingNotes.get(event.toolCallId);
 		if (!note) return;
@@ -222,10 +248,12 @@ export default function toolApprovalExtension(pi: ExtensionAPI): void {
 	});
 
 	pi.on("agent_end", () => {
+		batchReviews = undefined;
 		pendingNotes.clear();
 	});
 
 	pi.on("session_shutdown", (_event, ctx) => {
+		batchReviews = undefined;
 		pendingNotes.clear();
 		ctx.ui.setStatus(STATUS_KEY, undefined);
 	});
@@ -252,10 +280,49 @@ function autoApprovedMarker(value: unknown): AutoApprovedMarker | undefined {
 	return { toolName: record.toolName, provider: record.provider, model: record.model };
 }
 
-async function reviewToolRequest(
+async function reviewAssistantRequests(
 	ctx: ExtensionContext,
-	request: ToolApprovalRequest,
-): Promise<{ review: ToolReview; provider: string; model: string }> {
+	event: ToolCallEvent,
+	currentRequest: ToolApprovalRequest,
+): Promise<Map<string, CachedReview>> {
+	const reviews = new Map<string, CachedReview>();
+	const assistant = ctx.sessionManager
+		.getBranch()
+		.reverse()
+		.find((entry) => entry.type === "message" && entry.message.role === "assistant");
+	if (!assistant || assistant.type !== "message" || assistant.message.role !== "assistant") return reviews;
+	if (
+		!assistant.message.content.some(
+			(part) => part.type === "toolCall" && part.id === event.toolCallId && part.name === currentRequest.toolName,
+		)
+	)
+		return reviews;
+
+	const requests = assistant.message.content.flatMap((part) => {
+		if (part.type !== "toolCall" || (part.name !== "bash" && part.name !== "script_runner")) return [];
+		const request: ToolApprovalRequest = {
+			toolName: part.name,
+			input: part.id === event.toolCallId ? currentRequest.input : part.arguments,
+		};
+		if (request.toolName === "bash") {
+			const command = request.input.command;
+			if (typeof command !== "string" || !command.trim() || isAllowlistedBash(command)) return [];
+		}
+		return [{ toolCallId: part.id, request, requestJson: JSON.stringify(request) }];
+	});
+
+	for (let index = 0; index < requests.length && !ctx.signal?.aborted; index += MAX_CONCURRENT_REVIEWS) {
+		const group = requests.slice(index, index + MAX_CONCURRENT_REVIEWS);
+		const outcomes = await Promise.allSettled(group.map((item) => reviewToolRequest(ctx, item.request)));
+		for (const [offset, item] of group.entries()) {
+			const outcome = outcomes[offset];
+			if (outcome) reviews.set(item.toolCallId, { requestJson: item.requestJson, outcome });
+		}
+	}
+	return reviews;
+}
+
+async function reviewToolRequest(ctx: ExtensionContext, request: ToolApprovalRequest): Promise<ToolReviewResult> {
 	const requestJson = JSON.stringify(request);
 	const reviewer = REVIEW_MODELS.find((item) => item.provider === ctx.model?.provider);
 	const preferred = reviewer ? [reviewer] : [];

package/package.json

@@ -1,6 +1,6 @@
 {
 	"name": "@shanepadgett/tau-agent",
-	"version": "0.43.0",
+	"version": "0.44.0",
 	"description": "Tau is a custom agentic harness built with pi extensions",
 	"type": "module",
 	"main": "./src/index.ts",
@@ -35,7 +35,7 @@
 	],
 	"dependencies": {
 		"@ast-grep/wasm": "0.45.3",
-		"@shanepadgett/tau-tui": "0.43.0",
+		"@shanepadgett/tau-tui": "0.44.0",
 		"@vscode/tree-sitter-wasm": "0.3.1",
 		"image-size": "2.0.4",
 		"smol-toml": "1.8.0",