kenoma 1.1.0

A fake shell powered by raw LLM completion. The model hallucinates command output until it emits the next prompt.

kenoma

A shell powered by LLM completion.

Install

pip install kenoma

Or from source:

git clone https://github.com/a9lim/kenoma
cd kenoma
pip install -e .

For bitsandbytes quantization (CUDA only):

pip install kenoma[quantize]

Usage

kenoma                                # defaults to Qwen/Qwen2.5-0.5B
kenoma google/gemma-3n-E4B
kenoma /path/to/local/model

The model argument is any HuggingFace model id or a local path. This is meant for base/completion models, instruction-tuned models may not work properly.

Configuration

By precedence: CLI flags, then KENOMA_* environment variables, then a TOML config file at $XDG_CONFIG_HOME/kenoma/config.toml (falls back to ~/.config/kenoma/config.toml).

Example config:

model = "google/gemma-3n-E4B"
device = "auto"
temperature = 1.0
top_p = 0.95
repetition_penalty = 1.05
max_new_tokens = 2048
context_chars = 6000
history = 20
tmux_lines = 300
quantize = "none"
kv_cache = true
compile = false

The env var for any key is KENOMA_<KEY> uppercased, so KENOMA_MODEL=gpt2 kenoma works.

Flags:

• --version: print version and exit.
• --prompt: override the captured PS1. Multi-line prompts are not supported and fall back to a constructed user@host:cwd $ .
• --device {auto,cuda,mps,cpu}: auto resolves to cuda, then mps, then cpu.
• --quantize {none,4bit,8bit}: bitsandbytes quantization. Requires CUDA and the quantize extra.
• --no-kv-cache: disable KV cache reuse across turns.
• --compile: torch.compile the model with a static KV cache for faster decode (best on CUDA). The first turn pays a compile cost; cross-turn KV cache reuse is forfeited because the static cache doesn't expose crop().
• --history N: seed with the last N commands from shell history (0 disables).
• --tmux-lines N: if inside tmux, seed with the last N lines of pane scrollback (0 disables).
• --context-chars N: cap the rolling buffer at N chars.
• --max-new-tokens N: per-turn cap on generated tokens.

Cancelling a turn. Ctrl-C during generation cancels the current turn, invalidates the KV cache, and redraws the prompt. Ctrl-C at the input prompt exits.

License

AGPL-3.0-or-later. See LICENSE.