support tool approvals for pydantic-ai chatbot

frontend/src/components/chat/__tests__/chat-utils.test.ts

@@ -0,0 +1,269 @@
+/* Copyright 2026 Marimo. All rights reserved. */
+
+import type { UIMessage } from "ai";
+import { describe, expect, it } from "vitest";
+import { hasPendingToolCalls } from "../chat-utils";
+
+/**
+ * `hasPendingToolCalls` powers `sendAutomaticallyWhen` in `mo.ui.chat`:
+ * returns true only when the last assistant message *ends* with a tool
+ * call in a ready-to-round-trip state. Any trailing non-tool part (text,
+ * file, source-*, reasoning, data-*, new step-start) means the assistant
+ * has already answered and we leave the next turn to the user. The
+ * approval flow relies on this firing for `approval-responded`.
+ */
+
+const userMessage = (text: string): UIMessage => ({
+  id: `user-${text}`,
+  role: "user",
+  parts: [{ type: "text", text }],
+});
+
+const assistantToolMessage = (
+  parts: UIMessage["parts"],
+  id = "assistant-1",
+): UIMessage => ({
+  id,
+  role: "assistant",
+  parts,
+});
+
+describe("hasPendingToolCalls", () => {
+  it("returns false when there are no messages", () => {
+    expect(hasPendingToolCalls([])).toBe(false);
+  });
+
+  it("returns false when the last message is a user message", () => {
+    expect(hasPendingToolCalls([userMessage("hi")])).toBe(false);
+  });
+
+  it("returns false when the last assistant message has no tool parts", () => {
+    expect(
+      hasPendingToolCalls([
+        userMessage("hi"),
+        assistantToolMessage([{ type: "text", text: "hello!" }]),
+      ]),
+    ).toBe(false);
+  });
+
+  it("returns false while a tool is still streaming or awaiting approval", () => {
+    expect(
+      hasPendingToolCalls([
+        userMessage("delete it"),
+        assistantToolMessage([
+          {
+            type: "tool-delete_file",
+            toolCallId: "call-1",
+            state: "approval-requested",
+            input: { path: "secrets.env" },
+            approval: { id: "approval-1" },
+          } as unknown as UIMessage["parts"][number],
+        ]),
+      ]),
+    ).toBe(false);
+  });
+
+  it("returns true when the user has responded to an approval request", () => {
+    // The chat must auto-resume as soon as Approve/Deny is clicked.
+    expect(
+      hasPendingToolCalls([
+        userMessage("delete it"),
+        assistantToolMessage([
+          {
+            type: "tool-delete_file",
+            toolCallId: "call-1",
+            state: "approval-responded",
+            input: { path: "secrets.env" },
+            approval: { id: "approval-1", approved: true },
+          } as unknown as UIMessage["parts"][number],
+        ]),
+      ]),
+    ).toBe(true);
+  });
+
+  it("returns true when a tool reached a terminal output state", () => {
+    expect(
+      hasPendingToolCalls([
+        userMessage("run it"),
+        assistantToolMessage([
+          {
+            type: "tool-run_query",
+            toolCallId: "call-1",
+            state: "output-available",
+            input: { sql: "select 1" },
+            output: 1,
+          } as unknown as UIMessage["parts"][number],
+        ]),
+      ]),
+    ).toBe(true);
+  });
+
+  it("returns false when only some tool calls are ready", () => {
+    expect(
+      hasPendingToolCalls([
+        userMessage("two things"),
+        assistantToolMessage([
+          {
+            type: "tool-first",
+            toolCallId: "call-1",
+            state: "output-available",
+            input: {},
+            output: 1,
+          } as unknown as UIMessage["parts"][number],
+          {
+            type: "tool-second",
+            toolCallId: "call-2",
+            state: "input-available",
+            input: {},
+          } as unknown as UIMessage["parts"][number],
+        ]),
+      ]),
+    ).toBe(false);
+  });
+
+  it("returns false once the assistant has appended text after the tool result", () => {
+    expect(
+      hasPendingToolCalls([
+        userMessage("run it"),
+        assistantToolMessage([
+          {
+            type: "tool-run_query",
+            toolCallId: "call-1",
+            state: "output-available",
+            input: {},
+            output: 1,
+          } as unknown as UIMessage["parts"][number],
+          { type: "text", text: "The query returned 1." },
+        ]),
+      ]),
+    ).toBe(false);
+  });
+
+  it("returns false when a file part trails the completed tool call", () => {
+    // Regression: tool → text → file used to loop because only trailing
+    // text counted as "the assistant has answered".
+    expect(
+      hasPendingToolCalls([
+        userMessage("show me Starry Night"),
+        assistantToolMessage([
+          { type: "step-start" },
+          {
+            type: "tool-search_artwork",
+            toolCallId: "call-1",
+            state: "output-available",
+            input: { artist: "Van Gogh" },
+            output: { title: "The Starry Night" },
+          } as unknown as UIMessage["parts"][number],
+          { type: "text", text: "Here is the painting:" },
+          {
+            type: "file",
+            mediaType: "image/jpeg",
+            url: "https://example.com/starry-night.jpg",
+          } as unknown as UIMessage["parts"][number],
+        ]),
+      ]),
+    ).toBe(false);
+  });
+
+  it("returns false when a source-url part trails the completed tool call", () => {
+    expect(
+      hasPendingToolCalls([
+        userMessage("cite your sources"),
+        assistantToolMessage([
+          {
+            type: "tool-web_search",
+            toolCallId: "call-1",
+            state: "output-available",
+            input: { q: "marimo notebook" },
+            output: "found",
+          } as unknown as UIMessage["parts"][number],
+          { type: "text", text: "marimo is a reactive notebook." },
+          {
+            type: "source-url",
+            sourceId: "src-1",
+            url: "https://marimo.io",
+          } as unknown as UIMessage["parts"][number],
+        ]),
+      ]),
+    ).toBe(false);
+  });
+
+  it("returns false when a reasoning part trails the completed tool call", () => {
+    expect(
+      hasPendingToolCalls([
+        userMessage("explain"),
+        assistantToolMessage([
+          {
+            type: "tool-lookup",
+            toolCallId: "call-1",
+            state: "output-available",
+            input: {},
+            output: 1,
+          } as unknown as UIMessage["parts"][number],
+          {
+            type: "reasoning",
+            text: "Now I'll summarize.",
+          } as unknown as UIMessage["parts"][number],
+        ]),
+      ]),
+    ).toBe(false);
+  });
+
+  it("returns false when a new step-start follows the completed tool call", () => {
+    expect(
+      hasPendingToolCalls([
+        userMessage("multi-step"),
+        assistantToolMessage([
+          { type: "step-start" },
+          {
+            type: "tool-run_query",
+            toolCallId: "call-1",
+            state: "output-available",
+            input: {},
+            output: 1,
+          } as unknown as UIMessage["parts"][number],
+          { type: "step-start" },
+        ]),
+      ]),
+    ).toBe(false);
+  });
+
+  it("ignores providerExecuted tools", () => {
+    // Provider-side tools are resolved by the model, not the runtime, so
+    // they must not drive an auto-resume.
+    expect(
+      hasPendingToolCalls([
+        userMessage("hi"),
+        assistantToolMessage([
+          {
+            type: "tool-web_search",
+            toolCallId: "call-1",
+            state: "output-available",
+            input: {},
+            output: 1,
+            providerExecuted: true,
+          } as unknown as UIMessage["parts"][number],
+        ]),
+      ]),
+    ).toBe(false);
+  });
+
+  it("returns true for dynamic-tool parts in a terminal state", () => {
+    // `dynamic-tool` parts must drive auto-resume alongside `tool-*`.
+    expect(
+      hasPendingToolCalls([
+        userMessage("run it"),
+        assistantToolMessage([
+          {
+            type: "dynamic-tool",
+            toolName: "run_query",
+            toolCallId: "call-1",
+            state: "output-available",
+            input: {},
+            output: 1,
+          } as unknown as UIMessage["parts"][number],
+        ]),
+      ]),
+    ).toBe(true);
+  });
+});

frontend/src/components/chat/chat-utils.ts

@@ -5,7 +5,8 @@ import {
   type ChatAddToolOutputFunction,
   type FileUIPart,
   isToolUIPart,
-  type ToolUIPart,
+  lastAssistantMessageIsCompleteWithApprovalResponses,
+  lastAssistantMessageIsCompleteWithToolCalls,
   type UIMessage,
 } from "ai";
 import { useState } from "react";
@@ -17,7 +18,6 @@ import type {
   InvokeAiToolRequest,
   InvokeAiToolResponse,
 } from "@/core/network/types";
-import { logNever } from "@/utils/assertNever";
 import { blobToString } from "@/utils/fileToBase64";
 import { Logger } from "@/utils/Logger";
 import { getAICompletionBodyWithAttachments } from "../editor/ai/completion-utils";
@@ -169,69 +169,25 @@ export async function handleToolCall({
 }
 
 /**
- * Returns true if a tool call is "ready to be sent back to the server" — i.e.
- * either it has reached a terminal output state, or the user has just supplied
- * an approval response that the server hasn't seen yet.
- */
-function isToolCallReadyToSend(state: ToolUIPart["state"]): boolean {
-  switch (state) {
-    case "output-available":
-    case "output-error":
-    case "output-denied":
-    case "approval-responded":
-      return true;
-    case "input-streaming":
-    case "input-available":
-    case "approval-requested":
-      return false;
-    default:
-      logNever(state);
-      return false;
-  }
-}
-
-/**
- * Checks if we should send a message automatically based on the messages.
- * We auto-send when every tool call on the last assistant message has either
- * finished (output-available/error/denied) or has just received a user
- * approval response, and the assistant hasn't replied yet.
+ * Auto-send the next turn when the last assistant message ends with a
+ * tool call ready to round-trip. Any non-tool trailing part (text, file,
+ * source-*, reasoning, data-*, new step-start) means the assistant has
+ * already answered, so we leave the next turn to the user. State checks
+ * are delegated to the SDK to stay in sync with upstream.
  */
 export function hasPendingToolCalls(messages: UIMessage[]): boolean {
-  if (messages.length === 0) {
-    return false;
-  }
-
-  const lastMessage = messages[messages.length - 1];
-  const parts = lastMessage.parts;
-
-  if (parts.length === 0) {
-    return false;
-  }
-
-  // Only auto-send if the last message is an assistant message
-  // Because assistant messages are the ones that can have tool calls
-  if (lastMessage.role !== "assistant") {
+  const lastMessage = messages.at(-1);
+  if (!lastMessage || lastMessage.role !== "assistant") {
     return false;
   }
-
-  const toolParts = parts.filter(isToolUIPart);
-
-  if (toolParts.length === 0) {
+  const lastPart = lastMessage.parts.at(-1);
+  if (!lastPart || !isToolUIPart(lastPart)) {
     return false;
   }
-
-  const allToolCallsReady = toolParts.every((part) =>
-    isToolCallReadyToSend(part.state),
+  return (
+    lastAssistantMessageIsCompleteWithToolCalls({ messages }) ||
+    lastAssistantMessageIsCompleteWithApprovalResponses({ messages })
   );
-
-  // Check if the last part has any text content
-  const lastPart = parts[parts.length - 1];
-  const hasTextContent =
-    lastPart.type === "text" && lastPart.text?.trim().length > 0;
-
-  Logger.debug("All tool calls ready to send: %s", allToolCallsReady);
-
-  return allToolCallsReady && !hasTextContent;
 }
 
 export function useFileState() {

frontend/src/plugins/impl/chat/chat-ui.tsx

@@ -27,7 +27,10 @@ import {
 import React, { useEffect, useRef, useState } from "react";
 import { z } from "zod";
 import { renderUIMessage } from "@/components/chat/chat-display";
-import { convertToFileUIPart } from "@/components/chat/chat-utils";
+import {
+  convertToFileUIPart,
+  hasPendingToolCalls,
+} from "@/components/chat/chat-utils";
 import {
   type AdditionalCompletions,
   PromptInput,
@@ -186,7 +189,9 @@ export const Chatbot: React.FC<Props> = (props) => {
     error,
     regenerate,
     clearError,
+    addToolApprovalResponse,
   } = useChat({
+    sendAutomaticallyWhen: ({ messages }) => hasPendingToolCalls(messages),
     transport: new DefaultChatTransport({
       fetch: async (
         request: RequestInfo | URL,
@@ -440,6 +445,7 @@ export const Chatbot: React.FC<Props> = (props) => {
                   message,
                   isStreamingReasoning: status === "streaming",
                   isLast,
+                  addToolApprovalResponse,
                 })}
               </div>
               <div className="flex justify-end text-xs gap-2 invisible group-hover:visible">