Release Highlights

• Easier and more flexible configuration: Simplified config setup, moving away from a global CONFIG to use Config instances for more granular control. Typical use cases involve assigning different LLMs to different Roles. Please check out here for detailed usage.
• Introduced Context: Context is made up of config, repo, workspace, etc. Each Role can possess its own context, allowing more flexible context management.
• Added metagpt --init-config command: A more friendly setup process for pypi package users.
• Repo restructuring: Consolidated FileRepo and GitRepo responsibilities to simplify code and logic.
• Subscription to tags: Changed the concept of subscription to tags for improved readability and usability.
• Removal of explicit serialize/deserialize functions: Removed these functions from Role, used mixin instead for better clarity.
• Supported initializing ActionNode from pydantic.BaseModel: An easy and structured way to define ActionNode.
• WritePRD Action restructuring: Simplified code, enhanced effectiveness, and fixed buggy logic.
• Env restructuring (experimental): Re-abstracted Env to accommodate various environments more effectively; Uses mark_as_readable mark_as_writable to decorate APIs to uniformly observe and act on the environment through observe and step.
• Embraced Multi-modality: Added a vision agent based on GPT-4V, check out the example here!
• Improved Incremental development: Higher success rate of incremental development compared to previous version.
• Introduced CodeInterpreter: A general agent who solves problems through coding. The agent leverages on plans and tools, capable of tackling a wide range of problems, including data visualization, machine learning modeling, math reasoning, web scraping, webpage imitating, etc. We are adding examples here!

What's Changed

Backwards Incompatible Changes

• IMPORTANT: We removed config.yam, key.yaml and switched to a minimal config2.yaml. Please check the latest setup method here
• We removed Role._init_actions and added Role.set_actions, which allows actions setup at any time besides initialization. For Role initialization specifically, users should also use self.set_actions(...) now

Features

• Easier and more flexible configuration, introduced Context, embraced Multi-modality, improved Incremental development, introduced CodeInterpreter, and more (see release highlights) #852 by @geekan @mannaandpoem @garylin2099 @shenchucheng @voidking @iorisa @better629 @stellaHSR @orange-crow @buttogo @FromCSUZhou @luxiangtaoya
• Supported Generating class views from source code #676 by @iorisa

Bugfixes and improvements

• Improved testing SOP, increased test coverage, and fix bugs during testing, ensuring quality and preparing for more effective development. #673 #678 #685 #688 #689 #690 #694 #696 #698 #702 #703 #707 #709 #710 #711 #713 #715 #718 #727 #729 #863 #867 by @shenchucheng @voidking @iorisa @garylin2099
• Fixed is_human not working at Role initialization #715 by @kkdev163
• Fixed system_prompt error at BaseLLM #732 by @better629
• Constrain the language for the qa_engineer #743 by @arnaudgelas
• Fixed: requirements.txt not written to the disk #744 by @arnaudgelas
• Fixed multiple issues at qa_engineer.py, including crashing if no code_doc, generating tests for non-python files, trying to install dependencies when there are none #765 #775 by @arnaudgelas
• Upgraded Zhipu API #770 #861 by @better629
• Fixed prompt logic when defining to whom the message should be sent. #785 by @arnaudgelas
• Fixed JSONDecodeError when parsing LLM outputs #799 by @HuZixia
• Modified get_choice_delta_text for safer dict access while using ZHIPU API #801 by @pkdog