Refactor agent starter for multi-chat helper agents

[threepointone]

Summary

This PR refactors the starter from a single chat-agent example into a richer multi-chat launch point for building agent applications on Cloudflare.

The new architecture is centered around an Inbox root agent that owns app-level state and shared MCP configuration. Each conversation is represented by a Chat subagent, and focused helper agents such as Researcher and Planner are exposed through agentTool so the main assistant can delegate work while preserving inspectable helper transcripts.

What changed

• Reworked the server architecture around Inbox, Chat, Researcher, and Planner agents.
• Added multi-chat state and metadata management, including chat creation, rename, deletion, sorting, and last-message previews.
• Added shared MCP configuration at the Inbox level and proxied MCP tools into each Chat so external tools can be configured once and reused across conversations.
• Added retained helper-agent tools for research and planning, with guarded drill-in routing for helper transcripts.
• Removed the old data URL conversion workaround now that the AI SDK handles data URLs correctly.
• Removed aggressive default message pruning so the starter preserves full context by default; the README now documents pruning/summarization as production context-management options.
• Refactored the React UI into focused modules:
  • src/app.tsx for the app shell, header, inbox connection, and chat sidebar.
  • src/components/active-chat.tsx for the chat pane, composer, attachments, messages, and scheduling toasts.
  • src/components/mcp-panel.tsx for shared MCP server controls.
  • src/components/tool-views.tsx for tool cards, helper run cards, markdown output, and readonly drill-in panels.
  • src/constants.ts for shared starter constants.
• Improved loading behavior by moving Suspense boundaries down to the chat pane and drill-in panel, avoiding full-app loading flashes when switching chats or opening helper drill-ins.
• Rendered string tool/helper outputs with Streamdown so markdown responses from helper tools display correctly.
• Made helper drill-ins readonly to avoid confusing side conversations with the main chat.
• Added explanatory source comments for retained helper agents, shared MCP proxying, and subagent access guards.
• Updated wrangler.jsonc and regenerated env.d.ts for the root Inbox Durable Object binding.
• Rewrote the README to explain the multi-agent architecture, customization points, helper-agent workflow, shared MCP behavior, Durable Object routing, long-conversation guidance, model-provider swaps, and updated project structure.

Why

The starter is meant to be a launching point. The previous shape demonstrated a capable single chat agent, but it did not clearly show how to build a multi-chat app, delegate work to retained subagents, share MCP tools across conversations, or inspect helper-agent runs. This refactor keeps the demo approachable while giving users a more realistic foundation for production agent apps.

Test plan

• Ran npm run check successfully.
• Ran npx vite build successfully.
• Regenerated Cloudflare environment types with npm run types.
• Manually smoke test chat creation, switching, rename, and delete in the browser.
• Manually smoke test image attachment, timezone tool, approval tool, scheduling toast, helper-agent delegation, helper drill-in, and MCP server add/remove flows.

Notes for reviewers

• Inbox is the only directly bound Durable Object in wrangler.jsonc; Chat, Researcher, and Planner are reached as subagents through the Agents SDK routing layer.
• The helper registry is intentionally simple for a starter. Adding a new helper currently requires touching both server registration/guards and the UI helper display list.
• MCP discovery is still synchronous before each chat response. For production hardening, a follow-up could fall back to normal chat if MCP servers are slow or unavailable.

Made with Cursor