PicoBot

A multi-channel AI agent framework with a WebSocket gateway and TUI client, supporting OpenAI-compatible and Anthropic LLM providers, tool calling, session persistence, and cron-based scheduling.

System Architecture

graph TB
    subgraph Clients
        TUI["🖥️ CLI Chat (TUI)"]
        FS["📱 Feishu/Lark"]
    end

    subgraph Gateway["Gateway Server (127.0.0.1:19876)"]
        HTTP["HTTP Endpoints<br/>GET /health<br/>GET /ws (WebSocket upgrade)"]
        WS["WebSocket Handler"]
        CD["ChannelManager"]
        SP["SessionManager"]
        AL["AgentLoop"]
    end

    subgraph Bus["MessageBus"]
        IB["Inbound Channel"]
        OB["Outbound Channel"]
        CC["Control Channel"]
    end

    subgraph Storage
        SQLite[("SQLite<br/>.picobot_sessions.db")]
    end

    subgraph AI["AI Providers"]
        OpenAI["OpenAI / DashScope"]
        Anthropic["Anthropic Claude"]
    end

    TUI <-->|WebSocket| WS
    FS <-->|Webhook| HTTP

    CD -->|InboundMessage| IB
    IB -->|DialogEvent| SP
    CC -->|ControlMessage| SP
    SP <--> AL
    AL -->|API Call| OpenAI
    AL -->|API Call| Anthropic
    AL -->|Tool Call| Tools
    SP -->|OutboundMessage| OB
    OB --> CD
    SP --> SQLite
    Tools --> SQLite

    subgraph Tools
        Bash["Bash"]
        FileIO["File Read/Write/Edit"]
        Web["HTTP Request / Web Fetch"]
        Calc["Calculator"]
        Skill["Get Skill"]
        Msg["Send Message"]
        Cron["Cron Jobs"]
    end

Core Data Flow

sequenceDiagram
    participant Channel as Channel<br/>(CLI/Feishu)
    participant Bus as MessageBus
    participant SM as SessionManager
    participant AL as AgentLoop
    participant LLM as LLM Provider
    participant Tool as Tools

    Channel->>Bus: InboundMessage (user input)
    Bus->>SM: DialogEvent
    SM->>SM: Load/Resolve Session
    SM->>AL: Process (session state)
    AL->>LLM: ChatCompletionRequest
    LLM-->>AL: response / tool_calls
    alt Tool Calls
        AL->>Tool: execute tool
        Tool-->>AL: result
        AL->>LLM: continue with tool result
    end
    AL-->>SM: AgentProcessResult (text + token count)
    SM->>SM: Persist to SQLite
    SM->>Bus: OutboundMessage
    Bus->>Channel: response to user

Features

Multi-Channel Support

• CLI Chat Client — Full TUI with session management, Markdown rendering, slash commands
• Feishu (Lark) — Webhook-based integration with typing indicators and media support

Multi-Provider LLM

• OpenAI-compatible API (GPT-4, DashScope, Volcengine, etc.)
• Anthropic Messages API (Claude)
• Cross-provider JSON Schema normalization for tool calling compatibility

Session Management

• Multi-session conversations per channel/chat
• Create, switch, rename, archive, delete dialogs via slash commands or WebSocket
• SQLite-persisted session history with automatic TTL-based cleanup
• Context compression for long conversations approaching token limits

Tool System

Tool	Description
bash	Execute shell commands in workspace
file_read	Read file contents
file_write	Create/overwrite files
file_edit	Precise string substitution in files
http_request	Make HTTP API requests
web_fetch	Fetch and parse web pages
calculator	Evaluate mathematical expressions
get_skill	Load agent skills from local skill files
send_message	Send messages to other channels
cron_add/list/remove/enable/disable/update	Manage scheduled jobs

Scheduling

• Cron-based recurring jobs with optional timezone support
• One-shot (at) and interval (every) schedules
• Jobs trigger agent processing via specified channel/chat

Skills System

• Load Markdown skill files from ~/.picobot/skills and ~/.agent/skills
• Skills inject specialized system prompts for specific tasks
• Automatic hot-reload on file changes

Observability

• Observer pattern for agent and tool telemetry
• Events: AgentStart, AgentEnd, ToolCallStart, ToolCall
• Structured JSON logging with file rotation

Quick Start

Prerequisites

• Rust nightly (edition 2024) — use rustup to install

Build

cargo build

Configure

1. Create config.json (or ~/.picobot/config.json):

{
    "providers": {
        "openai": {
            "type": "openai",
            "base_url": "https://api.openai.com/v1",
            "api_key": "<OPENAI_API_KEY>"
        }
    },
    "models": {
        "gpt-4o": {
            "model_id": "gpt-4o",
            "temperature": 0.7,
            "max_tokens": 4096
        }
    },
    "agents": {
        "default": {
            "provider": "openai",
            "model": "gpt-4o",
            "max_tool_iterations": 20,
            "token_limit": 128000
        }
    }
}

2. Set API keys via .env file (one KEY=VALUE per line):

OPENAI_API_KEY=sk-xxxxx

Run

Start gateway server:

cargo run -- gateway

Binds 127.0.0.1:19876 by default. Override with --host and --port.

Connect CLI client:

cargo run -- chat

Connects to ws://127.0.0.1:19876/ws. Override with --gateway-url.

Configuration Reference

Config load order: ~/.picobot/config.json → ./config.json (fallback).

Full Config Structure

graph LR
    Config["config.json"]
    Config --> Providers["providers<br/>ProviderConfig{}"]
    Config --> Models["models<br/>ModelConfig{}"]
    Config --> Agents["agents<br/>AgentConfig{}"]
    Config --> Gateway["gateway<br/>GatewayConfig"]
    Config --> Client["client<br/>ClientConfig"]
    Config --> Channels["channels<br/>ChannelConfig{}"]
    Config --> Workspace["workspace_dir"]

    Providers --> PT["type (openai / anthropic)<br/>base_url<br/>api_key<br/>extra_headers"]
    Models --> MT["model_id<br/>temperature<br/>max_tokens"]
    Agents --> AT["provider (ref)<br/>model (ref)<br/>max_tool_iterations<br/>token_limit"]
    Gateway --> GT["host / port<br/>session_ttl_hours<br/>cleanup_interval_minutes<br/>session_db_path<br/>scheduler"]
    Channels --> CT["feishu: app_id, app_secret<br/>allow_from, agent, media_dir"]

Environment Variables

The .env file in the working directory is loaded manually (not via dotenv crate). Placeholders in config.json written as <VAR_NAME> are substituted at load time.

Gateway Config

Key	Type	Default	Description
host	string	127.0.0.1	Bind address
port	u16	19876	Listen port
session_ttl_hours	number	4	Inactive session expiration (hours)
cleanup_interval_minutes	number	60	Session cleanup interval
session_db_path	string	workspace .picobot_sessions.db	SQLite database path
scheduler.enabled	bool	false	Enable cron scheduler

Agent Config

Key	Type	Default	Description
provider	string	—	Provider name (key in providers)
model	string	—	Model name (key in models)
max_tool_iterations	number	20	Max tool call iterations per turn
token_limit	number	128000	Context window token limit

Slash Commands

Available in CLI chat and Feishu:

Command	Alias	Description
/new	/刷新	Create a new dialog
/list	/对话列表	List all dialogs
/switch <id>	—	Switch to a dialog
/rename <title>	—	Rename current dialog
/archive	—	Archive current dialog
/delete	—	Delete current dialog
/clear	/清空	Clear current dialog history

WebSocket Protocol

The gateway exposes a WebSocket endpoint at /ws. Messages use typed JSON with a type discriminator field.

Client → Server (WsInbound)

Type	Fields
user_input	content, channel?, chat_id?, sender_id?
create_session	title?
list_sessions	include_archived
load_session	session_id
rename_session	session_id?, title
archive_session	session_id?
delete_session	session_id?
clear_history	chat_id?, session_id?
get_slash_commands	—
ping	—

Server → Client (WsOutbound)

Type	Fields
assistant_response	session_id, response, tokens_used?, tool_calls?
session_list	sessions[]
session_loaded	session_id, messages[]
session_created	session_id, title
session_renamed	session_id, title
session_archived	session_id
session_deleted	session_id
slash_commands	commands[]
error	message
pong	—

HTTP Endpoints

Method	Path	Description
GET	/health	Health check — returns {"status":"ok","version":"x.y.z"}
GET	/ws	WebSocket upgrade for chat clients

Testing

# Unit tests (no external dependencies)
cargo test --lib

# Integration tests (require API keys)
cp tests/test.env.example tests/test.env
# Fill in your API keys in tests/test.env
cargo test --test test_integration -- --ignored
cargo test --test test_tool_calling -- --ignored
cargo test --test test_request_format -- --ignored

# Run all tests
cargo test -- --ignored

Integration tests are #[ignore] by default because they make real API calls.

Project Structure

├── src/
│   ├── main.rs              # CLI entrypoint (clap-based subcommands)
│   ├── lib.rs                # Module declarations
│   ├── gateway/              # HTTP/WS server, GatewayState initialization
│   │   ├── mod.rs
│   │   ├── http.rs           # Health endpoint
│   │   └── ws.rs             # WebSocket handler
│   ├── client/               # TUI chat client
│   │   ├── mod.rs
│   │   └── tui/              # Ratatui-based terminal UI
│   ├── channels/             # Channel integrations
│   │   ├── base.rs           # Channel trait
│   │   ├── cli_chat.rs       # CLI WebSocket channel
│   │   ├── feishu.rs         # Feishu/Lark webhook channel
│   │   ├── manager.rs        # ChannelManager
│   │   └── slash_command.rs  # Slash command parser
│   ├── bus/                  # Async message bus
│   │   ├── mod.rs            # MessageBus (tokio mpsc channels)
│   │   ├── message.rs        # Message types
│   │   └── dispatcher.rs     # OutboundDispatcher
│   ├── session/              # Session & dialog management
│   │   ├── mod.rs
│   │   ├── session.rs        # Session, SessionManager
│   │   ├── session_id.rs     # UnifiedSessionId
│   │   ├── commands.rs       # SessionCommand enum
│   │   └── events.rs         # SessionEvent, DialogInfo
│   ├── agent/                # LLM interaction loop
│   │   ├── mod.rs
│   │   ├── agent_loop.rs     # AgentLoop (stateless)
│   │   ├── context_compressor.rs  # Token estimation & summarization
│   │   └── system_prompt.rs  # System prompt builder
│   ├── providers/            # LLM API clients
│   │   ├── mod.rs            # Factory: create_provider()
│   │   ├── traits.rs         # LLMProvider trait
│   │   ├── openai.rs         # OpenAI-compatible client
│   │   └── anthropic.rs      # Anthropic Messages API client
│   ├── tools/                # Agent tools
│   │   ├── mod.rs            # create_default_tools()
│   │   ├── registry.rs       # ToolRegistry
│   │   ├── traits.rs         # Tool trait, ToolResult
│   │   ├── schema.rs         # Cross-provider JSON Schema cleaner
│   │   ├── bash.rs           # Shell command execution
│   │   ├── calculator.rs     # Math expression evaluator
│   │   ├── chat_manager.rs   # Session management tool
│   │   ├── cron.rs           # Cron job management tools
│   │   ├── file_read.rs      # File reader
│   │   ├── file_write.rs     # File writer
│   │   ├── file_edit.rs      # File editor (string substitution)
│   │   ├── get_skill.rs      # Skill loader tool
│   │   ├── http_request.rs   # HTTP request tool
│   │   ├── send_message.rs   # Cross-channel messaging
│   │   └── web_fetch.rs      # Web page fetcher
│   ├── skills/               # Skills loading from markdown files
│   │   └── mod.rs            # SkillsLoader, Skill
│   ├── storage/              # SQLite persistence
│   │   ├── mod.rs            # Storage, schema init
│   │   ├── session.rs        # Session CRUD operations
│   │   ├── message.rs        # Message persistence
│   │   ├── scheduler.rs      # ScheduledJob, JobRun storage
│   │   └── error.rs          # StorageError
│   ├── scheduler/            # Cron scheduler runtime
│   │   ├── mod.rs            # Scheduler, next_run_for_schedule()
│   │   └── types.rs          # Schedule enum (At/Every/Cron)
│   ├── observability/        # Telemetry observer pattern
│   │   └── mod.rs            # Observer trait, ObserverEvent, MultiObserver
│   ├── protocol.rs           # WebSocket message types (WsInbound/WsOutbound)
│   ├── config/               # Config loading & env substitution
│   │   └── mod.rs            # Config, LLMProviderConfig, load_env_file()
│   └── logging.rs            # Tracing subscriber init with file rotation
├── tests/
│   ├── test_integration.rs   # LLM provider integration tests
│   ├── test_tool_calling.rs  # Tool calling integration tests
│   ├── test_request_format.rs # Request format tests
│   ├── test_scheduler.rs     # Scheduler unit tests
│   ├── test.env.example      # Test environment template
│   └── test.env              # Actual test keys (gitignored)
├── reference/                # Third-party reference code (do not modify)
├── config.example.json       # Full config example
└── Cargo.toml

Key Dependencies

Crate	Purpose
axum + tokio-tungstenite	HTTP server & WebSocket
sqlx (SQLite)	Session/Message/Job persistence
reqwest (rustls)	LLM API & external HTTP calls
ratatui + crossterm	Terminal UI
clap	CLI argument parsing
tracing + tracing-subscriber	Structured logging
cron + chrono-tz	Cron schedule parsing
meval	Mathematical expression evaluation
uuid	Session/Dialog ID generation
dirs	Platform config directory resolution