Upgrading from 0.3

Version 0.4 adds the life portrait, assistant access and encrypted storage. The original fine-tuning workflow remains fully supported. The changes below affect users of 0.3.

Configuration file

The configuration version is now 0.4.01, and the command line warns at startup when your file doesn't match. Copy the template again and move your settings over one by one:

cp settings.jsonc settings.old.jsonc
cp settings.template.jsonc settings.jsonc

New sections:

Section	Purpose
security_args	Storage mode, see Security
codex_exec_args	Model, reasoning effort and concurrency when models are called through the Codex CLI
agent_distill_args	Model backend and windowing for portrait extraction

Other changes:

• QLoRA settings now go in train_sft_args.quantization, see Enable QLoRA.
• On older GPUs without bfloat16, set vllm_args.dtype instead of passing a command line flag.
• The examples/ directory has been removed. See Model Download for multimodal settings.

Every section is described in Configuration File.

Installation

The base install is now uv pip install -e ., which is enough for the life portrait and assistants. Fine-tuning still needs uv pip install --group main -e ., see Installation.

Server and port

weclone-cli server now serves the web UI and the APIs together, on port 5175 by default, without loading a model. Add --inference for the inference API, and keep the old port 8005 with:

weclone-cli server --inference --port 8005

Chat bots need a token

Bot platforms calling /v1 must use the token generated on the Bots page of the web UI as their API key. Any other key gets a 401. See Server and Bot Token.

weclone-cli test-model also connects to http://127.0.0.1:8005/v1 and asks for the shared password.

Storage mode and password

• Old configurations without security_args keep plaintext storage, so existing data and models work as before.
• The new template sets "storage_mode": "encrypted". In encrypted mode, old plaintext training data and models can't be used. Build the training data again from the original chats and retrain.
• The web UI asks you to set a shared password first. In encrypted mode, command line jobs unlock your data with the same password.