@kushuri12/ohiru

Provides a Telegram bot interface that receives user messages via Telegram's Bot API polling or webhook mechanism, routes them to an underlying LLM agent, and sends responses back through Telegram's message API. The agent maintains conversation context within Telegram chat sessions, enabling multi-turn dialogue without explicit session management by the user.

Manages stateful conversations by maintaining message history and context across multiple user interactions, passing accumulated context to an underlying LLM provider (OpenAI, Anthropic, or compatible API) for each new user message. The agent uses a prompt-based system to define behavior and instruction-following patterns, with context automatically appended to each API call.

Enables the agent to invoke external functions or APIs by leveraging the underlying LLM provider's function-calling capability (e.g., OpenAI's function calling, Anthropic's tool use). The agent receives function definitions, the LLM decides when to call them based on user intent, and results are fed back into the conversation context for the LLM to interpret and respond to.

Parses incoming Telegram messages to identify command patterns (e.g., /start, /help, /reset) and routes them to corresponding handler functions. Also handles callback queries from inline buttons, allowing structured user interactions beyond free-form text. The routing system decouples command handlers from the core agent logic, enabling modular command definitions.

Provides mechanisms to save, load, and reset conversation state (message history and context) for individual Telegram users or chats. State can be persisted to external storage (database, file system) or managed in-memory. Reset functionality clears conversation history, allowing users to start fresh conversations without restarting the bot.

Implements error handling for LLM API failures, Telegram API errors, and function call failures. When errors occur, the agent can gracefully degrade by returning error messages to users, retrying failed operations, or falling back to default responses. Error context is preserved for debugging and logging.

Implements rate limiting to prevent abuse of the Telegram bot and underlying LLM API. Can enforce per-user rate limits (e.g., max messages per minute), per-chat limits, or global limits. Quota tracking prevents excessive API costs by monitoring token usage or API call counts. When limits are exceeded, the agent can reject requests or queue them for later processing.

Provides built-in logging for agent operations including message routing, LLM API calls, function calls, and errors. Debug mode can be enabled to log detailed information about agent state, context, and decision-making. Logs can be output to console, files, or external logging services. Structured logging enables filtering and analysis of agent behavior.