llama_term

AI-assisted interactive shell for UNIX-like systems. Commands run in a persistent bash session; on failure (or via ASK()), an OpenAI-compatible model — or optional Anthropic / Gemini — proposes a fix. You confirm before any model-generated script runs.

Requirements

• Python 3.10+
• Linux / UNIX-like terminal (Parrot, Debian, Ubuntu, …)
• An OpenAI-compatible HTTP endpoint (Omniroute, Ollama /v1, LiteLLM, …) by default
• Python dependencies: pip install -r requirements.txt (includes optional packages defusedxml and lxml for DOCX support)

git clone https://github.com/Kevinscki/llama_term.git
cd llama_term
python3 -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt
cp .env.example .env   # set OPENAI_URL_API / ANTHROPIC_URL_API / GEMINI_API as needed
python3 llama_shell.py

Configuration

Variable	Meaning
API_TYPE	openai_http (default), anthropic_http, or gemini
OPENAI_MODEL / OPENAI_URL_ENDPOINT	Chat Completions URL (e.g. …/v1/chat/completions)
OPENAI_API / OPENAI_URL_API	True/False whether to send a bearer token; key when True
ANTHROPIC_MODEL / ANTHROPIC_URL_ENDPOINT	Anthropic Messages URL
ANTHROPIC_API / ANTHROPIC_URL_API	Same pattern for Anthropic x-api-key
GEMINI_API / GEMINI_MODEL	Cloud Gemini (set in .env, never commit)
MAX_INCLUDE_BYTES	Cap for INCLUDE() (default 200000)

Secrets come from the environment or .env (see .env.example). Rotate any key that was previously stored in source.

Privacy

• Default backend is OpenAI-compatible HTTP (often local Omniroute/Ollama).
• If API_TYPE is gemini or anthropic_http (or a remote OpenAI endpoint), included files and failure context may leave this machine. INCLUDE() warns when relevant.

Core commands

Command	Purpose
ASK() …	Ask the model; it generates bash from live context
FLAGS()	Toggle repeat / embed / exitcode context markers
INCLUDE() file	Attach file(s) for AI context (size + sensitive-path gates)
TOOL() NAME / TOOL() :N	Switch domain prompt by name or list position (DOCKER, PENTEST, …)
BUMP()	Clear chat turns; keep tool + live embeds
LOAD()	Warm the model
y / n / a	Confirm AI script — a = session auto-run for non-risky only
c / v / t	Copy script to clipboard · view it · run captured and attach STDOUT/STDERR to context
1-9	Multiple blocks: run just that block

Real TTY mode

The persistent shell is pipe-based so state survives every command, but TTY-hungry programs (vim, less, htop, ssh, python REPL, …) are wrapped in script(1) at run time — each gets a fresh PTY inside the persistent bash, so colors and full-screen UIs work while cd, env vars and functions still persist. Set LLAMA_REALTTY=auto|always|off (default auto).

Safety model

• Risky keyword hits always require an explicit y (session auto-run cannot bypass).
• Shell control trailers use a per-session random token (normal output cannot spoof them).
• Runtime files live under a per-process temp directory.

Known limitations

• TTY wrapping is per-command (script(1)); a wrapped command's output is relayed line-by-line, so exotic full-screen redraws can look slightly off.
• AI suggestions can be wrong — always review before y.
• See FIX.md for the repair backlog and remaining P2 work.