refactor: introduce write_todos tool for enhanced task management

- Added a new write_todos tool to facilitate the creation and management of planning lists within the chat interface.
- Updated system prompt with detailed instructions on using the write_todos tool, including usage patterns and restrictions.
- Enhanced the chat message handling to support the new tool, ensuring proper integration and user experience.
- Implemented UI components for displaying and interacting with the planning lists, including progress tracking and status indicators.
This commit is contained in:
Anish Sarkar 2025-12-26 14:37:23 +05:30
parent 1dd740bb23
commit eb70c055a4
14 changed files with 1138 additions and 1 deletions

View file

@ -0,0 +1,199 @@
"use client";
import { makeAssistantToolUI } from "@assistant-ui/react";
import { useAtomValue, useSetAtom } from "jotai";
import { Loader2 } from "lucide-react";
import { useEffect, useMemo } from "react";
import {
getCanonicalPlanTitle,
planStatesAtom,
registerPlanOwner,
updatePlanStateAtom,
} from "@/atoms/chat/plan-state.atom";
import { Plan, PlanErrorBoundary, parseSerializablePlan } from "./plan";
/**
* Tool arguments for write_todos
*/
interface WriteTodosArgs {
title?: string;
description?: string;
todos?: Array<{
id: string;
content: string;
status: "pending" | "in_progress" | "completed" | "cancelled";
}>;
}
/**
* Tool result for write_todos
*/
interface WriteTodosResult {
id: string;
title: string;
description?: string;
todos: Array<{
id: string;
label: string;
status: "pending" | "in_progress" | "completed" | "cancelled";
description?: string;
}>;
}
/**
* Loading state component
*/
function WriteTodosLoading() {
return (
<div className="my-4 w-full max-w-xl rounded-2xl border bg-card/60 px-5 py-4 shadow-sm">
<div className="flex items-center gap-3">
<Loader2 className="size-5 animate-spin text-primary" />
<span className="text-sm text-muted-foreground">Creating plan...</span>
</div>
</div>
);
}
/**
* Transform tool args to result format
* This handles the case where the LLM is streaming the tool call
*/
function transformArgsToResult(args: WriteTodosArgs): WriteTodosResult | null {
if (!args.todos || !Array.isArray(args.todos) || args.todos.length === 0) {
return null;
}
return {
id: `plan-${Date.now()}`,
title: args.title || "Planning Approach",
description: args.description,
todos: args.todos.map((todo, index) => ({
id: todo.id || `todo-${index}`,
label: todo.content || "Task",
status: todo.status || "pending",
})),
};
}
/**
* WriteTodos Tool UI Component
*
* Displays the agent's planning/todo list with a beautiful UI.
* Shows progress, status indicators, and expandable details.
*
* FIXED POSITION: When the same plan (by title) is updated multiple times,
* only the FIRST component renders. Subsequent updates just update the
* shared state, and the first component reads from it. This prevents
* layout shift when plans are updated.
*/
export const WriteTodosToolUI = makeAssistantToolUI<WriteTodosArgs, WriteTodosResult>({
toolName: "write_todos",
render: function WriteTodosUI({ args, result, status, toolCallId }) {
const updatePlanState = useSetAtom(updatePlanStateAtom);
const planStates = useAtomValue(planStatesAtom);
// Get the plan data (from result or args)
const planData = result || transformArgsToResult(args);
const rawTitle = planData?.title || args.title || "Planning Approach";
// SYNCHRONOUS ownership check - happens immediately, no race conditions
// ONE PLAN PER CONVERSATION: Only first write_todos call becomes owner
const isOwner = useMemo(() => {
return registerPlanOwner(rawTitle, toolCallId);
}, [rawTitle, toolCallId]);
// Get canonical title - always use the FIRST plan's title
// This ensures all updates go to the same plan state
const planTitle = useMemo(() => getCanonicalPlanTitle(rawTitle), [rawTitle]);
// Register/update the plan state - ALWAYS use canonical title
useEffect(() => {
if (planData) {
updatePlanState({
id: planData.id,
title: planTitle, // Use canonical title, not raw title
description: planData.description,
todos: planData.todos,
toolCallId,
});
}
}, [planData, planTitle, updatePlanState, toolCallId]);
// Update when result changes (for streaming updates)
useEffect(() => {
if (result) {
updatePlanState({
id: result.id,
title: planTitle, // Use canonical title, not raw title
description: result.description,
todos: result.todos,
toolCallId,
});
}
}, [result, planTitle, updatePlanState, toolCallId]);
// Get the current plan state (may be updated by other components)
const currentPlanState = planStates.get(planTitle);
// If we're NOT the owner, render nothing (the owner will render)
if (!isOwner) {
return null;
}
// Loading state - tool is still running
if (status.type === "running" || status.type === "requires-action") {
// Try to show partial results from args while streaming
const partialResult = transformArgsToResult(args);
if (partialResult) {
const plan = parseSerializablePlan(partialResult);
return (
<div className="my-4">
<PlanErrorBoundary>
<Plan {...plan} showProgress={true} />
</PlanErrorBoundary>
</div>
);
}
return <WriteTodosLoading />;
}
// Incomplete/cancelled state
if (status.type === "incomplete") {
if (status.reason === "cancelled") {
return null;
}
// For errors, try to show what we have from args or shared state
const fallbackResult = currentPlanState || transformArgsToResult(args);
if (fallbackResult) {
const plan = parseSerializablePlan(fallbackResult);
return (
<div className="my-4">
<PlanErrorBoundary>
<Plan {...plan} showProgress={true} />
</PlanErrorBoundary>
</div>
);
}
return null;
}
// Success - render the plan using the LATEST shared state
// This way, even if our result is old, we show the latest data
const planToRender = currentPlanState || result;
if (!planToRender) {
return <WriteTodosLoading />;
}
const plan = parseSerializablePlan(planToRender);
return (
<div className="my-4">
<PlanErrorBoundary>
<Plan {...plan} showProgress={true} />
</PlanErrorBoundary>
</div>
);
},
});
export type { WriteTodosArgs, WriteTodosResult };