Cellium Agent

Traditional Agents repeat mistakes, get stuck in loops, and never learn from experience. That's why we chose to make Agents evolve infinitely.

Based on microkernel architecture (EventBus + DI + BaseTool), supporting any OpenAI-compatible API.

Core design: Self-learning Agent driven by Control Loop, with adaptive decision optimization through Bayesian Bandit.

Thanks to the Strategy Gene research team. This project adopts their compact experience representation method, enabling the Agent to automatically learn avoidance strategies from failures.

What Can It Do

Cellium Agent is a versatile AI assistant that helps you with various tasks:

Feature	Example
File Operations	Read, write, search files, manage directory structures
Code Development	Write code, debug, refactor, explain code logic
Web Search	Search the internet for real-time information, summarize web content
Data Analysis	Process data files, generate reports and visualizations
Automation	Create scheduled tasks, e.g., "check weather every morning at 8 AM"
Component Extension	Create new components via natural language to extend Agent capabilities
Multi-Platform	Support WebUI, QQ, Telegram, Feishu (Lark), WeChat and other platforms

Usage Examples:

• "Read all files in the workspace directory and count the number of functions in each file"
• "Search for the latest AI technology developments, create a word component, and summarize into a report for me"
• "Create a scheduled task to check server status every hour, and notify me automatically when the status is abnormal"
• "Help me write a Python script to process CSV files and automatically generate data visualization charts"
• "Create a new component to monitor stock prices and notify me automatically when the price fluctuates by more than 5%"

Features

Feature	Description
Runtime Self-Awareness	Real-time perception of running state (progress, stagnation, loops, saturation), dynamically adjusting decisions
Control Loop Architecture	Closed-loop control of decision - execution - feedback - learning in each iteration
Self-Learning System	Action selection based on Bayesian Bandit, continuously optimizing decision strategies
Three-Layer Memory	Personality memory + Session memory + Long-term memory (FTS5 full-text retrieval + vector retrieval hybrid recall)
Heuristic Decision Engine	Rule-based feature extraction + Bandit for tie-breaking, balancing interpretability and learning ability
Tool Usage Control	Dynamic prohibition/recommendation of tool switching, avoiding loops from repeated tool calls
Sensitive Info Control	Auto-detect and redact sensitive info like private keys, tokens, passwords; supports write blocking
Component Hot-Plug	Files in app/components/ automatically load and take effect within 3 seconds
Component Sandbox Security	Three-layer protection: Process isolation + transparent path mapping + dangerous method interception
Event-Driven Architecture	Publish-subscribe pattern based on EventBus, loose coupling between components
Flash Mode	Skip memory injection to accelerate simple tasks
Multi-Channel Access	Support external platforms like QQ, Telegram, Feishu (Lark), WeChat, unified message routing, file transfer and injection through ChannelManager
Scheduled Tasks	Support interval tasks, daily tasks, weekly tasks. Create tasks via natural language, Agent executes and pushes results to the corresponding platform when triggered
Background Component Events	Components can run in background and actively trigger Agent execution, supporting real-time scenarios (e.g., crypto price monitoring, real-time data push, Agent auto-analysis)

Quick Start

One-Line Install (with Environment)

Windows

powershell -Command "Invoke-WebRequest -Uri 'https://github.com/Cellium-Project/Cellium-Agent/releases/latest/download/Cellium-Agent-Windows.zip' -OutFile 'Cellium-Agent-Windows.zip'; Expand-Archive -Path 'Cellium-Agent-Windows.zip' -DestinationPath '.' -Force; cd Cellium-Agent-Windows; .\CelliumAgent.exe"

Linux x64:

curl -LO https://github.com/Cellium-Project/Cellium-Agent/releases/latest/download/Cellium-Agent-Linux.tar.gz && tar -xzf Cellium-Agent-Linux.tar.gz && cd Cellium-Agent-Linux && ./start-cellium.sh

Linux ARM64:

curl -LO https://github.com/Cellium-Project/Cellium-Agent/releases/latest/download/Cellium-Agent-Linux-ARM64.tar.gz && tar -xzf Cellium-Agent-Linux-ARM64.tar.gz && cd Cellium-Agent-Linux-ARM64 && ./start-cellium.sh

macOS:

curl -LO https://github.com/Cellium-Project/Cellium-Agent/releases/latest/download/Cellium-Agent-macOS.tar.gz && tar -xzf Cellium-Agent-macOS.tar.gz && cd Cellium-Agent-macOS && ./start-cellium.sh

See INSTALL.md for more installation options

Install via pip (PyPI)

pip install cellium-agent

Launch the TUI interface after installation:

cellium

Launch the web service:

cellium-web

Optional: Feishu/Lark support requires lark-oapi:

pip install cellium-agent[lark]

Run from Source

pip install -r requirements.txt
python main.py

Main dependencies:

• FastAPI + Uvicorn (Web framework)
• PyYAML (Configuration parsing)
• DrissionPage (Web search and browser automation)
• websockets (QQ Bot WebSocket client)
• httpx (HTTP client for OpenAI-compatible API calls and external platform file upload)

Configure Models

Edit the config/agent/llm.yaml file to configure API keys, service addresses, and model names.

Start Service

python main.py

After startup, visit http://localhost:18000 to open the chat interface, and http://localhost:18000/docs to view API documentation. (Default port 18000, will automatically switch if occupied, check startup logs for actual port)

Core Architecture: Control Loop + Self-Learning

The core of Cellium Agent is a Control Loop driven decision-making system, combined with Bayesian Bandit for self-learning optimization.

┌─────────────────────────────────────────────────────────────────────────┐
│                           Learning Layer                                │
│  ┌─────────────┐    ┌─────────────┐    ┌─────────────────────────────┐  │
│  │   Policy    │    │  Bayesian   │    │      PolicyBanditMemory     │  │
│  │  Templates  │───▶│   Bandit    │◄───│  (Thompson Sampling Stats)  │  │
│  │             │    │             │    │                             │  │
│  └─────────────┘    └──────┬──────┘    └─────────────────────────────┘  │
│                            │                                            │
└────────────────────────────┼────────────────────────────────────────────┘
                             │ Select Policy
                             ▼
┌─────────────────────────────────────────────────────────────────────────┐
│                          Control Loop Layer                             │
│                                                                         │
│   ┌──────────┐     ┌──────────────┐     ┌──────────────┐               │
│   │   Step   │────▶│   Feature    │────▶│    Rule      │               │
│   │ (Start)  │     │  Extraction  │     │  Evaluation  │               │
│   └──────────┘     │              │     │              │               │
│        │           └──────────────┘     └──────┬───────┘               │
│        │                                        │                       │
│        │           ┌────────────────────────────┘                       │
│        │           ▼                                                    │
│        │     ┌──────────────┐     ┌──────────────┐                     │
│        │     │   Action     │◄────│   Action     │                     │
│        │     │  Candidates  │     │   Bandit     │                     │
│        │     │              │     │ (Tie-break)  │                     │
│        │     └──────┬───────┘     └──────────────┘                     │
│        │            │                                                  │
│        │            ▼                                                  │
│        │     ┌──────────────┐                                         │
│        │     │   Control    │                                         │
│        │     │   Decision   │                                         │
│        │     │   (Output)   │                                         │
│        │     └──────┬───────┘                                         │
│        │            │                                                  │
│        │     ┌──────┴───────┐     ┌──────────────┐                    │
│        └────▶│   Execute    │────▶│  End Round   │                    │
│              │              │     │              │                    │
│              └──────────────┘     └──────┬───────┘                    │
│                                          │                             │
│                                          ▼                             │
│                              ┌──────────────────────┐                  │
│                              │  Feedback Evaluator  │                  │
│                              │  (Feedback Eval)     │                  │
│                              │   - Segmented Eval   │                  │
│                              │   - n-step return    │                  │
│                              └──────────┬───────────┘                  │
│                                         │                              │
│                              ┌──────────┴───────────┐                  │
│                              ▼                      ▼                  │
│                    ┌─────────────────┐   ┌─────────────────┐          │
│                    │   Bandit Update │   │   Stats Persist │          │
│                    │   (Update Stats)│   │   (Persistence) │          │
│                    └─────────────────┘   └─────────────────┘          │
│                                                                         │
└─────────────────────────────────────────────────────────────────────────┘

Control Loop Workflow

Each loop contains 5 stages:

1. Feature Extraction

   • Heuristic engine extracts current state features
   • Includes: stagnation iterations, progress trends, repetition scores, context saturation, etc.

2. Rule Evaluation

   • Hard rules provide action candidate sets
   • Example: detect loops and candidate [redirect, compress]

3. Bandit Tie-break (Action Selection)

   • When candidate actions > 1, Bandit intervenes
   • Uses Thompson Sampling + Heuristic Bias to select optimal action

4. Execute & Feedback

   • Execute selected action (continue/retry/redirect/compress/terminate)
   • FeedbackEvaluator performs segmented evaluation of this round's performance

5. Learning & Update

   • Use n-step return to accumulate rewards
   • Update Bandit's Beta distribution parameters
   • Regularly decay old data to prevent overfitting

PEOP Loop (Plan-Execute-Observe-RePlan Loop)

The PEOP Loop is an extension module of the Control Loop, implementing an adaptive plan-execute cycle. This module dynamically adjusts strategy based on task complexity: simple tasks receive direct responses, complex tasks automatically enable multi-step planning; during execution, results are continuously validated, and local replanning is triggered when deviations are detected, achieving efficient and reliable task decomposition through explicit state management:

┌─────────────────────────────────────────────────────────────┐
│              Plan-Execute Engine State Machine               │
│                                                              │
│  ┌─────────┐    ┌─────────┐    ┌─────────┐                  │
│  │ OBSERVE │───▶│  PLAN   │───▶│ EXECUTE │◄──────────────┐  │
│  │ Observe │    │  Plan   │    │ Execute │  Validation    │  │
│  └─────────┘    └─────────┘    └────┬────┘  Success → Next │  │
│       ▲                             │                      │  │
│       │                   Validation│                      │  │
│       │                      Failed │                      │  │
│       │                             ▼                      │  │
│       │                        ┌─────────┐  Replan Success │  │
│       │                        │ REPLAN  │─────────────────┘  │
│       │                        │ Replan  │                   │  │
│       │                        └────┬────┘                   │  │
│       │                             │                      │  │
│       └─────────────────────────────┘  Max Replan Reached   │  │
│                                        or Task Complete     │  │
│                                        ▼                    │  │
│                                      ┌─────┐                │  │
│                                      │DONE │                │  │
│                                      │Done │                │  │
│                                      └─────┘                │  │
└─────────────────────────────────────────────────────────────┘

Core Mechanisms:

Mechanism	Description
Batch Planning	Generate multi-step execution plan at once (1-5 steps), reducing LLM call frequency
State-Driven	5-phase explicit state machine (OBSERVE/PLAN/EXECUTE/REPLAN/DONE)
In-Execution Validation	Automatic result validation after each step (semantic matching + Jaccard similarity + purpose-driven)
Local Replanning	Preserve successful steps on validation failure, only replan failed and subsequent steps

Workflow:

1. OBSERVE: Analyze user input, understand task goals and context
2. PLAN: LLM generates structured plan, each step contains: tool name, parameters, execution purpose, expected result
3. EXECUTE: Execute plan steps sequentially, automatic validation after each step
   • Validation Success → Continue to next step
   • Validation Failed → Enter REPLAN phase
4. REPLAN: Preserve successful steps, only regenerate plan for failed and subsequent steps
5. DONE: All steps executed successfully, or max replanning reached

Design Characteristics:

• Efficient: Multi-step plan generated once, zero LLM calls during execution phase
• Reliable: Expectation validation uses semantic matching, avoiding misjudgment (e.g., "function X" and "get_X" considered matching)
• Stable: Local replanning avoids total overhaul, maintaining context continuity
• Observable: 5-phase state machine provides clear execution trace for debugging and monitoring
• Collaborative: State information synchronized to Control Loop in real-time, replanning triggers redirect decision

Configuration:

• max_plan_steps=5: Maximum 5 steps per plan
• max_replans=3: Maximum 3 replanning attempts

Action Types & Strategies

Code definition: ACTION_TYPES = ["continue", "retry", "redirect", "compress", "terminate"]

Action	Description	Heuristic Bias Condition
continue	Continue current direction	Progress score > 0.5 or stagnation iterations = 0
retry	Maintain direction but correct strategy	Mild stagnation (1 <= stuck < threshold) or progress trend 0~0.3
redirect	Change direction/tool	Repetition score > 0.5 or stagnation >= stuck_threshold
compress	Compress context	Context saturation > 0.6 or stagnation >= stuck_threshold // 2
terminate	Terminate session	Hard rule triggered: output loop and exact_repetition_count >= 5

Self-Learning Mechanism

Policy - Bandit - Action Three-Layer Architecture:

┌─────────────────────────────────────────┐
│           Policy Templates              │
│  ┌─────────┬───────────┬─────────────┐  │
│  │ default │ efficient │ aggressive  │  │
│  │(stuck=3)│ (stuck=2) │  (stuck=5)  │  │
│  └────┬────┴─────┬─────┴──────┬──────┘  │
│       │          │            │         │
│       ▼          ▼            ▼         │
│  ┌─────────────────────────────────┐    │
│  │      Bayesian Bandit            │    │
│  │  Thompson Sampling selects      │    │
│  │  optimal Policy                 │    │
│  │  - Sample from Beta dist        │    │
│  │  - Select highest expected      │    │
│  │    return Policy                │    │
│  └─────────────┬───────────────────┘    │
│                │                        │
│                ▼                        │
│  ┌─────────────────────────────────┐    │
│  │        Action Bandit            │    │
│  │  Tie-break within candidates    │    │
│  │  - Heuristic provides bias      │    │
│  │  - Dynamic threshold adjustment │    │
│  └─────────────────────────────────┘    │
└─────────────────────────────────────────┘

Learning Process:

1. Policy Selection: At session start, Bayesian Bandit selects current optimal policy from multiple Policies (default/efficient/aggressive)
2. Threshold Injection: Selected Policy parameters (e.g., stuck_iterations=3) inject into HeuristicEngine and ActionBandit
3. Action Learning: After each round, update Action's Beta distribution based on FeedbackEvaluator's score
4. n-step return: Accumulate rewards from recent n rounds, supporting delayed feedback and sequence optimization
5. Data Decay: Decay old data every 50 sessions (decay factor 0.99) to prevent overfitting

Feedback Evaluation

Uses segmented design, first distinguishing success/failure, then optimizing details:

• Success Branch: Base score 1.0, deduct efficiency and cost

  • Iteration penalty: fewer iterations = higher score
  • Token penalty: deduct points if exceeding threshold
  • Smoothness reward: bonus for no stagnation

• Failure Branch: Base score 0.0, deduct points based on stagnation degree

  • More stagnation iterations = more points deducted
  • Error type affects deduction magnitude

Microkernel Architecture

┌─────────────────────────────────────────────────────────────┐
│                        EventBus                              │
│              (Publish-Subscribe, Loose Coupling)            │
├─────────────────────────────────────────────────────────────┤
│  ┌──────────┐  ┌──────────┐  ┌──────────┐  ┌───────────┐  │
│  │  LLM     │  │  Memory  │  │  Tools   │  │ Heuristics│  │
│  │  Engine  │  │  System  │  │          │  │  Engine   │  │
│  └──────────┘  └──────────┘  └──────────┘  └───────────┘  │
├─────────────────────────────────────────────────────────────┤
│                     AgentLoop (Main Loop)                  │
│  ┌────────────┐  ┌────────────┐  ┌────────────────────┐    │
│  │   Control  │  │   Tool     │  │   Prompt           │    │
│  │   Loop     │  │  Executor  │  │   Context Builder  │    │
│  └────────────┘  └────────────┘  └────────────────────┘    │
└─────────────────────────────────────────────────────────────┘

Core Components

Component	Description
AgentLoop	Event-driven core main loop, coordinating LLM, tools, memory
LLM Engine	Unified LLM interface, built-in 40+ model registry, auto-detecting context window, tool support, max output
ThreeLayerMemory	Three-layer memory: Personality + Session + FTS5 + optional vector retrieval
HeuristicEngine	Heuristic rule engine, serves as feature extractor for Bandit
ControlLoop	Control loop core, decision - execution - feedback - learning each round
ActionBandit	Action selector, Thompson Sampling + Heuristic Bias
LearningIntegration	Learning module integration, Policy selection and parameter injection
EventBus	Event bus, loose coupling communication between components
BaseTool	Tool base class, declarative command registration pattern

Component Sandbox Security

Agent-written component code runs in an isolated sandbox environment with three-layer security protection:

Layer	Mechanism	Function
Process Isolation	Components run in separate subprocess	Crash won't affect main process, communicate via Queue
Transparent Path Mapping	File operation paths mapped to sandbox_root	Prevent path traversal, protect real system files
Dangerous Method Interception	Runtime interception of os.system/subprocess	Block system command execution

Component code can use open(), os.listdir() normally, but paths are transparently mapped to the sandbox_root/ folder in the project directory. Meanwhile, dangerous methods like os.system(), subprocess.run() are intercepted at runtime.

Agent Runtime Self-State Awareness

Agent perceives running state in real-time through FeatureExtractor, dynamically adjusting decision strategies:

State Dimension	Feature	Description
Progress	progress_score	Task completion progress estimate (0-1)
	stuck_iterations	Consecutive non-progress iterations
	is_making_progress	Whether progress is being made
Trend	progress_trend	EMA-smoothed progress trend (-1 to 1)
	convergence_rate	Convergence speed
	is_plateau	Whether in plateau phase
Tool	unique_tools_used	Number of different tools used
	tool_diversity_score	Tool diversity score
	repetition_score	Tool repetition call score
	pattern_detected	Detected loop pattern
Context	context_saturation	Context saturation (0-1)
	context_saturation_level	Saturation level: idle/normal/warn/redirect/stop
Quality	error_rate	Tool call error rate
	empty_result_rate	Empty result rate
	result_quality_score	Comprehensive result quality score
Output	exact_repetition_count	LLM output exact repetition count
	is_output_loop	Whether in output loop

Adaptive Adjustment Mechanisms:

1. Stagnation Detection: Trigger redirect/retry when stuck_iterations > threshold
2. Context Pressure Awareness: Trigger compress when saturation > 0.7, stop when > 0.95
3. Tool Loop Detection: Trigger redirect to change tool when repetition_score > 0.5
4. Output Loop Detection: Force terminate when exact_repetition_count >= 5
5. Dynamic HardConstraint: Real-time generation of control commands (e.g., REDIRECT/COMPRESS/RETRY) based on state

Three-Layer Memory

Layer	Implementation	Description
Personality	personality.md	Static personality setting file
Session	MemoryManager	Short-term context, automatically maintains bounded history
Long-term	FTS5 + Repository	Vector retrieval + hybrid recall, supports knowledge extraction and archiving

Vector Retrieval:

Long-term memory supports optional API-based vector embedding for semantic search:

• Dimensions: Configurable (default 1536 for text-embedding-3-small)
• Provider: Any OpenAI-compatible embedding API
• Similarity: Cosine similarity
• Hybrid Recall: FTS5 full-text retrieval + vector similarity fusion sorting
• Configuration: config/agent/embedding.yaml — enable/disable, set API key, model, base URL
• On-Demand: Disabled by default, no external dependencies required when off

Structured Schema & Category System:

Long-term memory uses a hierarchical classification design. Schema types determine data structure, while Categories determine content types:

Schema Types (4 types):

Schema Type	Purpose	Included Categories
general	General memory, session notes	general, user_info, command, project
profile	User profile	preference
project	Project-related	project
issue	Troubleshooting	troubleshooting, code

Category Classification (7 types):

Category	Description	Source
general	Daily conversations, Q&A	General memory
user_info	User preferences, session goals	Session notes (goal/goal_history)
command	Executed operation commands	Session notes (completed)
project	Project config, key findings	General memory / Session notes (finding)
preference	User profile information	General memory
troubleshooting	Error records, solutions	General memory / Session notes (error)
code	Code-related records	General memory / Session notes

Session Compression Auto-Extraction:

During session compression, the following note types are automatically extracted and mapped to corresponding Categories for long-term memory storage:

• goal / goal_history → user_info
• completed → command / code
• finding → project / code / command
• error → troubleshooting
• pending → general

Sensitive Information Control:

The memory system has built-in sensitive information detection and protection mechanisms:

• Auto-Detection: Identifies sensitive content like private keys, AWS keys, GitHub tokens, API keys
• Redaction: Automatically replaces sensitive values with [REDACTED]
• Write Blocking: High-risk sensitive information (e.g., private keys) is blocked from memory by default
• Classification Marking: Sensitive memories are tagged and can be filtered out during retrieval

Heuristic Decision Rules

Rule	Description
MaxIterationRule	Iteration count limit termination
TokenBudgetRule	Token budget exhaustion termination
EmptyResultChainRule	Empty result chain detection
NoProgressRule	No progress detection (EMA smoothing)
SameToolRepetitionRule	Same tool + parameter repeated call detection
PatternLoopRule	Pattern loop detection
ParameterSimilarityRule	Parameter similarity detection

Component System

Feature	Description
Auto Discovery	Scan components/ directory, auto-discover component classes inheriting BaseCell
Hot-Plug	ComponentWatcher background monitoring, 3-second interval scanning, dynamic load/unload
Tool Wrapping	ComponentToolRegistry wraps BaseCell as BaseTool, injects into AgentLoop
Trust Whitelist	Components require user /trust confirmation, persisted to trusted_components.json
Config Auto-Maintenance	Auto-append discovered new components to settings.yaml enabled_components

Reserved Tool Names (cannot be overridden by components):

• shell — ShellTool
• memory — MemoryTool

QQBot File Transfer Component (qq_files)

Supports file transfer between QQBot and local:

Command	Function	Example
download	Download file from QQ	{"url": "...", "filename": "doc.pdf"}
send_file	Send file to QQ	{"target_id": "...", "file_path": "...", "is_group": false}
send_image	Send image to QQ	{"target_id": "...", "image_path": "..."}
list_downloads	List downloaded files	{}

File Save Path: workspace/downloads/qq/

Native Browser Operation Component (web_fetch)

Headless browser component based on DrissionPage, supports web automation:

Command	Function	Example
navigate	Visit specified URL	{"url": "https://example.com"}
get_screenshot	Screenshot (supports element-level)	{"full_page": false, "selector": "#content"}
find_qrcode	Find page QR code	{}
js_action	Page action (click/input/scroll_to)	{"selector": "button", "action": "click"}
find_button	Find button element	{"value": "Login"}
get_page_info	Get page info	{}
scroll	Scroll page	{"direction": "down", "amount": 500}
save_cookies / load_cookies	Cookie persistence	{"path": "cookies.json"}

Use Cases:

• Web content crawling and analysis
• Automated login flow (supports QR code recognition)
• Web screenshots and visual verification
• Form auto-fill and submission

Screenshot Save Path: workspace/web_fetch_screenshots/domain_timestamp.png

Core Mechanisms:

• _cell_registry: cell_name → ICell instance
• get_all_tools(): AgentLoop dynamically reads component tools at runtime
• get_tool_definitions(): Returns LLM-format tool definitions

Tool Operation Visibility:

• All tool calls generate user-friendly operation descriptions
• Supports _intent parameter for custom description (highest priority)
• Example: {"command": "read", "path": "test.py", "_intent": "Reading configuration file"}

Skill Management System (skill_installer + skill_manager)

Complete Skill package management solution, supporting installation from archives, list display, search, detail view, and uninstallation.

Installation Methods:

• Supports .zip, .tar.gz, .tgz, .tar archive formats
• Archive must contain SKILL.md file
• Auto-parses YAML Frontmatter metadata from SKILL.md

Management Functions:

Component	Function	Command/Interface
skill_installer	Install Skill	install_from_archive(path)
skill_installer	Refresh Index	refresh_index()
skill_manager	List all Skills	list(show_details=True)
skill_manager	Search Skills	search(query)
skill_manager	Get Details	get_info(name)
skill_manager	Uninstall Skill	uninstall(name)

Frontend Interface:

• Settings page provides Skill management panel
• Supports archive upload installation
• Supports filtering by name, description, category
• Detail modal displays complete metadata

Directory Convention:

components/
├── skill_installer.py        # Skill installer
├── skill_manager.py          # Skill manager
└── skills/                   # Skill installation directory
    ├── skill-a/              # Installed Skill
    │   └── SKILL.md
    ├── skill-b/
    │   └── SKILL.md
    └── _index.json           # Index file (auto-generated)

Scheduled Tasks (scheduler)

Three task types are supported:

Type	Description	Example
Interval Task	Execute at fixed intervals	Remind to drink water every 30 minutes
Daily Task	Execute at a fixed time every day	Send daily report at 9 AM
Weekly Task	Execute at a fixed time every week	Send weekly report at 10 AM every Monday

Usage: Create scheduled tasks via natural language. When triggered, Agent automatically executes the task content. Supports complex task scenarios:

• "Check server status and report results every hour"
• "Query today's weather at 8 AM every day"
• "Summarize this week's work progress every Friday at 5 PM"

Task execution results are automatically pushed to the platform where the task was created (WebUI/QQ/Telegram/Feishu/WeChat).

Background Component Event Trigger

Components can run in the background and actively trigger Agent execution, enabling real-time data push and auto-analysis.

Real-time Scenario Examples:

• Crypto price monitoring: Push data when price breaks threshold, Agent auto-analyzes trends
• Server status monitoring: Push alerts when CPU/memory is abnormal, Agent generates diagnostic reports
• Real-time news subscription: Push when keyword news detected, Agent auto-summarizes key points
• Database change listening: Push when critical data changes, Agent executes corresponding logic

Event trigger results are automatically pushed to the corresponding session (WebUI/QQ/Telegram/Feishu/WeChat).

Strategy Gene (GEP) Integration

This project adopts the Strategy Gene design proposed in the paper "From Procedural Skills to Strategy Genes: Towards Experience-Driven Test-Time Evolution" (arXiv:2604.15097).

Implementation

The paper proposes encoding experience into compact Gene objects (~230 tokens) instead of documentation-heavy Skill packages (~2,500 tokens). This project implements this in the Control Loop's Hard Constraint layer:

• Task Matching: Match Gene templates via LLM intent classification on user input
• Dynamic Injection: Inject matched Gene as runtime state into LLM (task-level dynamic injection, does not pollute conversation history)
• Experience Evolution: Automatically extract Avoid_Cues from failure feedback and update Gene

Implemented Features

Paper Concept	This Project Implementation
Gene Structure	[HARD CONSTRAINTS] + [CONTROL ACTION] + [AVOID]
Task Matching	TaskSignalMatcher LLM intent matching
Avoid_Cues	Auto-extract from failure feedback, write to [AVOID] section
Version Management	version field + evolution_history tracking changes
Effect Evaluation	success_rate, avg_reward, consecutive_success/failure
Cross Combination	GeneComposer merges multiple Genes for multi-task scenarios

Star History

License

Apache 2.0