AI agent (sudo agent) — in-console, no PTY

Rconsole ships a native, line-mode AI coding agent so you get tool-using AI assistance without fighting a browser-hostile PTY TUI. It calls an OpenAI-compatible (or Anthropic) API directly and can run tools inside your current target (host or VM — see vm mode) — bash, read_file, write_file, list_dir — all rendered through the normal line console.

Providers (first match wins)

Keys are read from your sudo env store (persisted per user, injected into every session) or the server process environment. Setting them needs no redeploy:

sudo env set OPENCODE_API_KEY=oc-...        # from https://opencode.ai/auth
sudo env set OPENCODE_MODEL=hy3-free        # any Zen /v1/chat/completions model
sudo agent

OpenRouter / local OpenAI-compatible servers work by setting OPENAI_BASE_URL (and OPENAI_MODEL). OpenCode Zen note: the GPT-5.x family on Zen requires the separate /v1/responses API (not yet wrapped here); use a chat/completions model such as hy3-free, deepseek-v4-flash, glm-5.2, kimi-k3, or nemotron-3.5-lightning-free.

Usage

sudo agent                       # interactive session (chat persists across commands)
sudo agent "write a flask app and run it"   # one-shot task, then stays open
sudo agent models                # list models from the current provider
sudo agent help                  # show agent slash commands

Inside a session, slash commands are available:

/help        show help            /clear      forget the conversation
/model <n>   switch model         /provider   show provider/model/base
/keys        list keys (masked)   /models     list provider models
/compact     summarise context    /system     print system prompt
/exit        leave the agent

Robustness

• Automatic retry/backoff on 429/5xx and connection errors (pooled requests.Session).
• Persistent conversation across sudo agent invocations (stored per tab; cleared on /clear or /exit).
• Context compaction — older turns are summarised automatically when the transcript grows large, to cap token usage and cost.
• Token usage report per turn and per session.
• Safety guard — on the host, the bash tool refuses obviously destructive commands (rm -rf /, mkfs, dd on disks, shutdown, …); run those directly in the console if you truly intend them. Inside your VM the guard is off because those are your own files (not because PRoot isolates the host), and vm reset rebuilds the system.

Edit this page on GitHub