@kevin666aa I will take a look at this some time later today, thanks for posting this PR!👌

The test fails.

@kevin666aa Let me know if this is still in draft or if it's ready for review

The test fails.

in the build workflow. why pip uninstall -y openai? It fails because openai is needed.

I am currently testing this PR, it works for a bit but I eventually get this error.

    assert isinstance(compressed_message, str), f"compressed_message should be a string: {compressed_message}"
AssertionError: compressed_message should be a string: {
  "role": "assistant",
  "content": null,
  "function_call": {
    "name": "subprocess",
    "arguments": "{\n  \"command\": \"nx generate @nrwl/react:component CampaignList --project=my-app\"\n}"
  }
}

I have the following function defined in my code.

def run_cli_command(command):
    try:
        # Run the CLI command in a separate process
        process = subprocess.Popen(
            command, shell=True, stdout=subprocess.PIPE, stderr=subprocess.PIPE, text=True)

        # Wait for the process to complete and capture the output
        stdout, stderr = process.communicate()

        # Check the return code to see if the command was successful
        return_code = process.returncode
        if return_code == 0:
            print("Command executed successfully.")
            print("Standard Output:")
            print(stdout)
        else:
            print("Command failed with the following error:")
            print(stderr)

    except Exception as e:
        print("An error occurred:", e)

   "functions": [
        {
            "name": "subprocess",
            "description": "run nx command in subprocess.",
            "parameters": {
                "type": "object",
                "properties": {
                    "command": {
                        "type": "string",
                        "description": "Valid nx command to execute.",
                    }
                },
                "required": ["command"],
            },
        },
    ],

[<OpenAIObject at 0x10f833c50> JSON: {
  "index": 0,
  "message": {
    "role": "assistant",
    "content": "SYSTEM:\nCompressed Content of Previous Chat:\n\nUSER: I made a mistake with the previous command and want to go back.\n\nASSISTANT: None\n\nUSER: Make sure to provide the \"command\" argument when calling the subprocess function.\n\nASSISTANT: None\n\nUSER: Oops, I made a mistake with the previous command. I want to go back.\n\nASSISTANT: None\n\nUSER: using nx nextjs generator create a component for the \"hive.ai\" project\n\nASSISTANT: None\n\nFUNCTION_CALL: subprocess - Execute \"nx g @nrwl/next:component hive-ai\"\n\nASSISTANT: None\n\nFUNCTION_CALL: subprocess - Execute \"nx g @nrwl/next:component hive-ai\"\n\nFUNCTION_RETURN (from \"func_name: subprocess\"): None\n\nFUNCTION_RETURN (from \"func_name: subprocess\"): None\n\nASSISTANT: Provide command to create component for \"hive.ai\" using Nx Next.js generator\n\nUSER: Provide command to create component for \"hive.ai\" using Nx Next.js generator\n\nASSISTANT: use \"nx g @nrwl/next:component --project=hive.ai\"\n\nUSER: use \"nx g @nrwl/next:component --project=hive.ai\"\n\nASSISTANT: None\n\nFUNCTION_CALL: subprocess - Execute \"nx g @nrwl/next:component --project=hive.ai\"\n\nASSISTANT: None\n\nFUNCTION_CALL: subprocess - Execute \"nx g @nrwl/next:component --project=hive.ai\"\n\nFUNCTION_RETURN (from \"func_name: subprocess\"): None\n\nFUNCTION_RETURN (from \"func_name: subprocess\"): None",
    "function_call": {
      "name": "subprocess",
      "arguments": "{\n  \"command\": \"nx g @nrwl/next:component hive-ai\"\n}"
    }
  },
  "finish_reason": "function_call"
}]

@AaronWard @Siafu @sonichi I made several changes to the compression agent. I fix several bugs and revise the prompt. It would be great if you can try out the compression notebook and maybe test with your own examples?

@Siafu Let me know if this fixes your issue.

Does this PR affect the default behavior of the ConversableAgent? If so, a wide-scale regression test is needed.
In general, any change to the core functionality of ConversableAgent needs to go through scrutiny and has a chance of not getting merged.
If the goal is to get merged fast, please consider alternative options.

Does this PR affect the default behavior of the ConversableAgent? If so, a wide-scale regression test is needed. In general, any change to the core functionality of ConversableAgent needs to go through scrutiny and has a chance of not getting merged. If the goal is to get merged fast, please consider alternative options.

This PR will affect the default behavior at very minimal level: The newly registered on_token_limit will terminate the conversation if it has found the token limit exceeds the max_token of the LLM. Basically catch the ratelimit before call API.
This can also be set configurable: the on_token_limit will be registered only when the user wants it.

Other than this, it will not affect default behavior at all. The compression is enabled only if the user wants to use the compression feature and passes in compress_config.

Is there an option of not altering ConversableAgent? It will lower the bar of merging. Otherwise, I really have to ask you for a long list of tests and it will take me a long time to finish the review.

@AaronWard @Siafu @sonichi I made several changes to the compression agent. I fix several bugs and revise the prompt. It would be great if you can try out the compression notebook and maybe test with your own examples?

@Siafu Let me know if this fixes your issue.

testing now

@kevin666aa Thanks for the nice PR. The compression feature is very important and asked by many users. But this PR involves changes of logic in ConversableAgent, the consequences of which are unclear yet. To be safe and to merge this PR sooner, I will suggest creating a new class by inheriting theConversableAgent instead of directly revising it. E.g., You can create a CompressibleAgent in the contrib folder. This also allows the users to start experimenting with this once it is merged. Once we later confirm it indeed does not affect the existing utility of the ConverableAgent and the compression functionality is well received, we can integrate it into the ConversableAgent.

But this PR involves changes of logic in ConversableAgent, the consequences of which are unclear yet.

@qingyun-wu this should be apparent with rigid and comprehensive unit testing. The impact of changes to the code should be measurable if there is any hope for longevity for any project. I can begin writing some tests but in all fairness ConversableAgent is pretty large so this should be a group effort

@qingyun-wu @AaronWard I agree with both of you. We should add tests, and before we have the tests in place, let's be cautious when modifying the core. BTW @afourney is working on tests too.

Thanks, then I will close this PR and add a compressible agent in the contrib first