v0.3.1

Highlights

• Secure sudo execution: the agent now runs sudo -v interactively (password entered directly into sudo), then executes sudo -n … with captured output so the password never reaches Python memory while the LLM still receives command output.
• Configurable System Prompt: override via environment variables (CLI_AGENT_SYSTEM_PROMPT_FILE / CLI_AGENT_SYSTEM_PROMPT), persistent settings (~/.cliagent/settings.json), or CLI (--system-prompt-file). Priority: env file > env text > config file > config text > default prompt.
• Iteration limit UX: when the maximum number of iterations is reached, the agent now asks whether to continue or adjust the task instead of stopping abruptly.

Technical Notes

• Version bumped to 0.3.1.
• Tests: uv run -q pytest → 44 passed.

0.3.0 – Windows support, shell mode, config & history

Highlights:

Windows support (PowerShell / CMD detection)
New shell execution mode with fallback to LLM
Unified config directory ~/.cliagent
Persistent command history (readline / pyreadline3 on Windows)
Improved error detection & natural language fallback
Cross-platform clear screen
Refactored providers layout & system context reporting
Details:

Added modules: agent.config, agent.core_agent, agent.utils, interface.shell_interface, input_handler package
Conditional dependency: pyreadline3 on Windows
Removed legacy flat scripts (agent.py, utils.py, input_handler.py)
Breaking:

CLI internals and module paths changed; upgrade requires reinstall.
Install / Upgrade:

uv tool install --force "git+https://github.com/anfedoro/cli-agent.git@v0.3.0"

Quick start/use:

cli-agent [-h] [--verbose] [--provider {openai,gemini,lmstudio}] [--model MODEL] [--no-reasoning] [--mode {chat,shell}] [--trace]
               [--no-restore]

LLM Terminal Agent

options:
  -h, --help            show this help message and exit
  --verbose, -v         Show detailed token usage information
  --provider {openai,gemini,lmstudio}, -p {openai,gemini,lmstudio}
                        LLM provider to use (default: openai)
  --model MODEL, -m MODEL
                        Model to use for the selected provider
  --no-reasoning        Disable reasoning process for faster responses (LM Studio)
  --mode {chat,shell}   Run mode: chat (conversation) or shell (command mode) (default: shell)
  --trace, -t           Enable trace mode (show LLM execution details, only in shell mode)
  --no-restore          Don't restore initial directory on exit (shell mode only)

License: MIT

Initial experimental release (chat-only, 3 AI providers)

First experimental cut.

Features:

Single interactive chat mode
Providers: OpenAI (gpt-*), Google Gemini, LM Studio (local)
Not included:

Windows support
Shell / exec mode
Config files or history
Advanced error handling
Status: Experimental / unstable. Interfaces may change without notice. Use at your own risk.

Install:

uv tool install "git+https://github.com/anfedoro/cli-agent.git@v0.2.0"

Quick start:

usage: cli-agent [-h] [--verbose] [--provider {openai,gemini,lmstudio}] [--model MODEL] [--no-reasoning]

LLM Terminal Agent

options:
  -h, --help            show this help message and exit
  --verbose, -v         Show detailed token usage information
  --provider {openai,gemini,lmstudio}, -p {openai,gemini,lmstudio}
                        LLM provider to use (default: openai)
  --model MODEL, -m MODEL
                        Model to use for the selected provider
  --no-reasoning        Disable reasoning process for faster responses (LM Studio)

Requirements:

Python 3.12+
Relevant API key (except for local LM Studio)
License: MIT