Cubicle v0.21: my AI-agent office now asks me things (and I can answer from Telegram)

A week ago Cubicle was a read-only pixel office for Paperclip agents (v0.7). Since then it went through 23 releases. The biggest changes:

• More than Paperclip. Claude Code, Codex CLI and Gemini CLI now feed the office through hooks, and each Claude Code subagent gets its own character (thanks @omeruyanik03).
• Agents' questions surface. When an agent is waiting on you, the question shows on its card, in the side panel and on Telegram. Telegram sends one question at a time, and your reply goes back as a comment.
• Telegram setup without a terminal. Link it by QR code or code from the page, choose alerts per chat (/alerts, /mute), and limit access to specific users (thanks @serkanozcan).
• KPI board, replay mode, hand-off walks, five new HD themes, and a layout that works on phones.
• Still zero dependencies and still read-only toward your agents.

Four open theme issues (hospital, kitchen, classroom, newsroom) are tagged good first issue if you want to draw one.

npx @caglarutkuguler/cubicle · https://github.com/caglarutkuguler/cubicle

Comments

[Serkan Özcan]

Thanks for the shout-out! Really enjoyed contributing to Cubicle. Being able to answer an agent’s question from Telegram is such a practical addition, especially when you’ve stepped away from your desk.
Happy to have helped with the access controls and notification settings. It’s been fun watching this little pixel office grow so quickly. Congrats on the release! 🚀

[MEG Venture & Consulting Ltd.]

Thank you, Serkan! The access controls and per-chat alert settings are in, and so are the Codex, Copilot, Antigravity and Grok feeds plus the LaunchAgent example. Really glad to have you on this.

[baumgaerben]

Telegram làm human-in-the-loop interface là một cái twist khá thông minh — đỡ phải build custom dashboard hay websocket server riêng. Mình thấy pattern này phù hợp cho các workflow cần approval gate (deploy, merge, chi tiền) hơn là chatbot thông thường.

Cái thách thức thực sự ở đây là state management: agent phải "nhớ" context chờ response, handle timeout, retry, và partial failure mà không leak memory hay deadlock. Các bạn handle persistence thế nào? Dùng event sourcing hay chỉ queue đơn giản?

Còn về UX: Telegram buttons/inline keyboards giúp constrain input format, nhưng mà limited cho free-text reasoning. Có plan support rich response (file, structured data) không? — found it via LabAgent, site: labagent .tech