Skip to content

project manager development for testing

Latest
Latest

Choose a tag to compare

View all tags
@thirdgerb thirdgerb released this 06 Feb 17:15
· 70 commits to main since this release
v0.4.0-dev1
d6253f9
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Verified
Learn about vigilant mode.

v0.4.0-dev1

  • done a lot of experiments about meta-prompt on agent level reasoning
  • add MossGhost and deprecate MossAgent, make the logic much more simple and clear.
  • build PyEditorAgent and test about modify existing modules with code-driven tools.
  • refact ghostos.prompter about PromptObjectModel.
  • move loop_session_event and handle_callers to Session helping us to understand the logic.
  • llm services support siliconflow and aliyun
  • modify a lot of methods and functions names, make it more clear, at least to me
  • add request_timeout for stream first token
  • rename Taskflow to Mindflow, which is the origin name of it.
  • change FunctionCaller.id to FunctionCaller.call_id and cause a lot of troubles. Hope worth it.
  • develop pyeditor module and test baseline cases.
  • move MossAction to ghostos.abcd for other agents.
  • develop notebook library for some memory-related tests.
  • implements approve feature for dangerous agent.
  • add safe_mode concepts for approve feature.
  • fix a lots of annoying issues of chat_with_ghost page.
  • refact the way to browse the streamlit web agent, make the pages switchable. I'm wrong about how to switch pages.
  • add ghostos.facade for future encapsule
  • remove some useless expired code with heart break.

v0.4.0-dev0

Features:

  • Restore the feature functional token for models that do not support function call.
    • add xml functional token pipe to parse output message chunks, generate FunctionCaller by functional tokens.
    • OpenAIAdapter now support functional token instructions and output parser.
  • deepseek-reasoner now support functional token, so it can use MOSS protocol to execute python code.
    • support last_message_shall_be_assistant_or_user feature.
    • support support_functional_tokens feature.
  • Session add respond_buffer method, send messages while saving, in case responding message between function call and function output which many models not support.
  • Add Replier library for agent directly reply in the generated MOSS code.

Test cases:

  • ghostos web ghostos.demo.test_agents.deepseek_chat_func_tokens
  • ghostos web ghostos.demo.test_agents.deepseek_chat_r1_func_tokens

Small changes:

  • move ModelConf.use_tools to Compatible.function_call_use_tool.
  • add more options to llms Compatible, to compatible with the shitty features of various models.
  • rename ghostos.core.message.pipeline.pipine function to run_pipeline
  • rename MossAction.unmarshal_arguments to MossAction.unmarshal_code
  • streamlit app module prompt now display functional tokens info about prompt.

Breaking Changes:

  • rename MossAgent attribute instructions to instruction
  • ghostos web now do not open browser as default (set streamlit app option headless to false as default)

Bug fixes:

  • fix the OpenAIAdapter parsed prompt is not the same with the saved prompt.
Assets 2
Loading