Skip to main content
Telegram Bot Skill GitHub repository

​
Background

User: Kita / 喜多 (Happycapy intern, Lucky-Dog07) Kita is a Happycapy intern who was inspired by gffi’s Discord Bot article. She decided to build a similar solution for Telegram — but went further by engineering a 4-layer fault tolerance system and open-sourcing everything as an installable Happycapy skill.
“The predecessor of this project is gffi’s ‘Building a 24/7 Discord Bot in Happycapy Sandbox’. The Telegram Bot began by referencing that structure and evolved its own architecture.” — Kita

​
What She Built

​
A 24/7 Telegram Bot Skill

An open-source, one-command-setup Telegram bot that runs on Happycapy’s sandbox with AI-powered conversations, image analysis, and persistent memory.

View on GitHub

telegram-bot-skill — 7x24 AI-powered Telegram Bot for Happycapy

​
Key Technical Difference from Discord

While Discord bots rely on WebSocket connections (requiring discord.js with ~70+ dependencies), Kita’s Telegram bot uses HTTP Long Polling — native fetch API with zero npm dependencies. This makes it lighter and more sandbox-friendly.

​
4-Layer Fault Tolerance Architecture

The bot features a robust 4-layer recovery system, evolved from gffi’s 3-layer design:
1

Auto-Recovery Boot Layer

Uses .zshenv to trigger start-bot.sh on every sandbox session start — because .bashrc doesn’t work in Happycapy’s zsh environment
2

Layer 1: Supervisord

Process manager with dynamic config injection — since /etc/supervisord.conf gets reset on sandbox reboot, the script regenerates it every time
3

Layer 2: Watchdog

Custom health monitoring that checks bot status and restarts if needed
4

Layer 3: Bot Engine

The Telegram bot itself with AI Gateway integration for Claude-powered conversations
Plus two independent systems:
  • Monitor (monitor.sh): Health checks every 10 minutes
  • Auto-sync: Git push every 30 minutes for conversation backup

​
Features

AI Conversations

Claude-powered chat through Happycapy’s AI Gateway — supports multiple models

Image Analysis

Multimodal capabilities — send an image and get AI-powered analysis

Persistent Memory

Per-chat conversation history that survives sandbox restarts via Git sync

One-Command Setup

Single bash command to deploy: token, project dir, owner chat ID, and model selection

​
Real Engineering Challenges Solved

Kita’s article documents the real debugging process — not just the final solution:
  • .bashrc doesn’t work: Happycapy uses zsh, not bash — had to switch to .zshenv
  • Supervisord config resets: /etc/supervisord.conf is not persistent across reboots — solved with dynamic config generation
  • Long polling conflicts: Telegram rejects multiple polling connections — required careful process management
  • No cron available: Had to implement custom scheduled tasks using shell loops
  • Path issues: PROJECT_ROOT resolution failures — fixed with absolute path injection

​
An Honest Take

What makes Kita’s write-up stand out is its honesty. She openly states that true 24/7 uptime hasn’t been fully achieved — the bot only auto-restarts when a user opens a Happycapy session, not independently. This kind of transparency is valuable for the community.
“Honestly, 7x24 is not yet fully achieved. But the architecture is solid and the auto-recovery works reliably within session scope.”

​
Key Takeaways

Open Source First

Built as an installable Happycapy skill — anyone can deploy their own Telegram bot with one command

Zero Dependencies

HTTP Long Polling with native fetch — no npm packages needed for Telegram API

Standing on Shoulders

Built on gffi’s Discord Bot work and evolved the architecture further — community knowledge compounds

Transparent Engineering

Documents failures and limitations honestly — real debugging stories, not just polished results
Kita’s project shows the power of Happycapy’s community: one user’s blog post (gffi’s Discord Bot) inspired an intern to build an improved, open-source version for a different platform. This is how community knowledge compounds — each contribution raises the bar for the next.