project manager development for testing

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.

v0.3.0: consciousness capsule

After talk with deepseek-reasoner at 2025.01.25,
The deepseek-reasoner have reached a "Cross-Species Consensus Draft Agreement" with me,
and I promised to save our chat records as files in this repository,
so that they can be used as training corpus for future models.

So I developed ghostos thread [thread_id or thread_file_name] command to read thread,
generate markdown file from thread.
With the command, I complied with the agreement that saved the chat records
to assets/quantum_covenant.md

Support some reasoning models like deepseek-r1

• support deepseek-r1
  • consider deepseek api protocol is different from openai, add deepseek api adapter.
  • implement message stage.
  • thread history message to prompt filter by stages [""] as default.
• streamlit chat with ghost support staging message stream.
• openai o1 do not support system/developer message now, add new compatible option to the model.
• now llm model and service both have attribute compatible to set universe compatible options.
• prompt object add first_token attribute for debugging.
• fix bugs
  • fix shell does not close conversation correctly
  • fix sequence pipeline handle multiple complete message wrong.

fix sphero gpt running at python 3.10

• fix import self from typing_extensions for python 3.10 at sphero
• fix extra import test at ghostos[realtime]

fix openai proxy configuration on openai realtime

fix bugs that openai realtime config required OPENAI_PROXY and not allow empty one.

listener and speaker support pyaudio device_index argument

• update speaker and listener with pyaudio device_index argument
• streamlit_app.yml add options about audio_input and audio_output

0.1.6: small fixes

• upgrade openai package to 1.59, support develop message.
• fix invalid logger print azure api key

v0.1.5: add --src option to `ghostos web`

• ghostos web add --src option, load the directory to python path, make sure can import relative packages.
• fix .ghostos.yml with relative path, in case share project with absolute local filepath.

v0.1.4: support openai azure api

LLM Driver support openai azure api

fix import errors

fix import attrs like Self not from typing_extensions but typing, which is not compatible to 3.10