GolemBot

GolemBot integrates with several specialized Coding Agent CLIs. Each engine has unique invocation methods, output formats, and session management characteristics. The supported engines are:

• Cursor Agent CLI: Uses the agent binary. Supports stream-json and includes a --stream-partial-output flag for deduplication.
• Claude Code CLI: Uses the claude binary. Supports stream-json and handles mixed assistant/user events.
• OpenCode CLI: Uses the opencode binary. Uses NDJSON parts and offers an HTTP Server API option.
• Codex CLI: Uses the codex binary. Uses NDJSON and requires ChatGPT OAuth. Includes a --skip-git-repo-check flag.

GolemBot includes four core built-in skills that are automatically initialized during golembot init or golembot onboard. These skills define the fundamental capabilities of an assistant, such as personal assistance, IM optimization, multi-bot coordination, and proactive messaging.

Available Built-in Skills:

• general: A personal assistant for tasks like file management, script execution, and information retrieval.
• im-adapter: Optimizes response formatting and tone for instant messaging platforms (Slack, Discord, Telegram, etc.).
• multi-bot: Enables coordination and peer awareness when multiple bots run in the same fleet.
• message-push: Allows the agent to proactively send messages to channels or users via API.

This demo showcases how multiple GolemBot instances can collaborate within a shared group channel using a mock group adapter. It demonstrates several advanced multi-agent patterns:

• Group-scoped sessions: All bots share a single session key per room, identified by channelType:chatId.
• smart mode: A bot (e.g., researcher) observes every message but only speaks when it has something valuable to add. If it has nothing to contribute, it outputs [PASS] to signal silence.
• mention-only mode: A bot (e.g., coder) remains silent unless it is explicitly @mentioned in a message.
• Multi-bot collaboration: Bots can @mention each other to hand off tasks.
• maxTurns protection: A safety mechanism to prevent infinite loops in bot-to-bot conversations.
• Group memory: Bots maintain persistent context using files located at memory/groups/<room>.md.

The Data Analyst Template relies on the following directory structure to function:

• data/: Contains the source data files (CSV, JSON, etc.) to be analyzed.
• reports/: The destination directory where the assistant saves generated analysis reports.
• skills/data-analysis/calc.py: A Python helper script used by the agent to perform statistical calculations.

The Code Reviewer Template consists of the following directory structure:

• reviews/: The output directory where the generated review reports are stored.
• skills/code-review/: Contains the definition for the code-review Skill used by the assistant.

The Content Operations Assistant template uses the following directory and file structure to manage its operations:

• content/: Stores content drafts.
• reports/: Stores operations briefings.
• schedule.md: Manages the content publishing schedule.
• competitors.md: Used for competitor monitoring.

The Customer Support Bot Template consists of the following key files and directories:

• faq.md: The primary knowledge base containing your FAQ content in Q&A format.
• unanswered.md: An auto-generated log that tracks questions the bot was unable to answer.
• skills/faq-support/: The directory containing the specific Skill definition used to power the customer support logic.

GolemBot's interaction model on Discord depends on the chat type and mention status:

Key Technical Details

• Sessions: Each DM conversation and each server channel maintains its own independent session context.
• Markdown: Discord natively supports Markdown. GolemBot sends responses as-is, and Discord renders formatting (bold, italic, code blocks, etc.) automatically.
• Message Limits: If an AI response exceeds 2,000 characters, Discord messages are automatically split into chunks.

To isolate conversation history for different users, pass a sessionKey to the chat() method.

• Different keys: Sessions run concurrently.
• Same keys: Sessions are serialized (one must finish before the next starts).

// Two users can chat simultaneously
const task1 = assistant.chat('Task A', { sessionKey: 'user-alice' });
const task2 = assistant.chat('Task B', { sessionKey: 'user-bob' });

The demo uses a MockGroupAdapter which utilizes a shared Node.js EventEmitter (named GroupRoom) as a channel bus. Any adapter instance that joins the same room emitter will receive all messages and can broadcast replies.

To move from a mock environment to a production environment, you should replace the MockGroupAdapter with a platform-specific adapter such as FeishuAdapter or SlackAdapter while maintaining the ability to join and broadcast to a shared room.

The general skill transforms the agent into a personal AI assistant.

Capabilities:

• Answering questions, brainstorming, and providing suggestions.
• File operations: Reading and writing files to organize notes or manage tasks.
• Automation: Executing scripts and commands.
• Information retrieval and summarization.

Persistent Memory: This skill uses a notes.md convention for cross-session memory. The agent reads this file at the start of a conversation and writes to it when important information is identified.

Skill Management: The agent can autonomously search for and install community skills from ClawHub. If a user requests a capability the agent lacks, it will suggest relevant skills from ClawHub. All commands support the --json flag for structured output.