AIR Studio

AI-augmented R analysis environment — a fork of Posit RStudio Server (AGPLv3) with a built-in AI coding agent and per-user conda R-environment switching.

AIR Studio installs standalone next to any existing RStudio Server on the same host: its own port (7878), paths (/usr/lib/air-studio, /etc/air-studio, ~/.local/share/air-studio), service (air-studio-server.service), PAM profile, and auth cookies. Nothing collides.

Features over stock RStudio Server

• Console AI agent — press Shift+Tab in the console to toggle agent mode. The prompt and input turn green; agent replies render with a light-blue background; R commands the agent executes use the normal console styling. Ships as a self-contained Node bundle (no npm install at runtime).
• Conda R-environment switcher — toolbar dropdown lists public environments (e.g. /opt/miniforge3/envs/*) and each user's personal environments (~/{miniconda3,miniforge3,anaconda3,.conda}/envs/*, filtered per user). Switching restarts the session in the selected R; the choice persists per user. A systemd timer rescans every 5 minutes.
• AIR Studio branding — product title, tab title, and toolbar logo.

Install

sudo dpkg -i air-studio-server_<ver>_amd64.deb air-studio-extras_<ver>_amd64.deb

air-studio-extras depends on the exact air-studio-server version (the chat protocol is version-matched); install both from the same release.

Configure the LLM gateway (operator default, optional)

The admin can set a server-wide default provider via any OpenAI-compatible endpoint:

sudo systemctl edit air-studio-server

[Service]
Environment=BIOAGENT_BASE_URL=http://your-gateway:8000/v1
Environment=BIOAGENT_API_KEY=your-key
Environment=BIOAGENT_MODEL=your-model

sudo systemctl restart air-studio-server

Optional tuning: BIOAGENT_CONTEXT_TOKENS (default 131072), BIOAGENT_OUTPUT_TOKENS (default 32768). Works with vLLM, OneAPI, OpenRouter, DeepSeek, and any /v1/chat/completions endpoint.

Per-user providers (console commands)

Each user can manage their own providers in agent mode (Shift+Tab) — no admin needed, takes precedence over the server default:

addprovider("http://your-gateway:8000/v1", "api-key")   # add; auto-lists & activates first model
showmodel("your-gateway")                               # list models ([x] = active)
setmodel("your-gateway", "model-id")                    # pick model & activate

Bare addprovider() / showmodel() print the current state. Registry: ~/.config/air-studio/llm-providers.json (0600). Without any provider configured, the agent answers with setup guidance instead of failing.

Context management

The agent tracks context usage per conversation and compacts automatically:

• ≥ 60% of the context window: a one-line notice suggests /compact
• ≥ 80%: history is folded into an LLM-generated summary before your next message, automatically
• manual: type /compact in agent mode at any time; the summary replaces the history and is shown in the console

Thresholds are tunable via BIOAGENT_SUGGEST_PCT / BIOAGENT_AUTO_COMPACT_PCT (fractions, default 0.6 / 0.8); window size via BIOAGENT_CONTEXT_TOKENS (default 131072).

Interrupting, queueing, and steering

While the agent is working on a turn:

• Enter — submits your message to the queue; it runs when the current turn finishes (messages run in arrival order)
• Shift+Enter — steers: interrupts the current turn immediately and submits your message right away; already-queued messages keep their places and run after it
• Esc Esc (double-press within 500 ms) — interrupts the current turn and discards queued messages; the partial reply is kept in the conversation

Interrupting also stops any in-flight R execution started by the agent. The backend acknowledges each action inline ([queued] …, [interrupted — N queued message(s) discarded]).

Building from source

git clone git@github.com:yuumeHuang/AIRstudio.git
cd AIRstudio
RSTUDIO_VERSION_MAJOR=2026 RSTUDIO_VERSION_MINOR=05 RSTUDIO_VERSION_PATCH=1 \
RSTUDIO_VERSION_SUFFIX='+225' RSTUDIO_BOOST_REQUESTED_VERSION=1.91.0 \
RSTUDIO_NODE_VERSION=22.22.2 \
./package/linux/make-package Server DEB

The extras package (agent backend + conda machinery) builds from the same tree: conda-r-versions/build-extras.sh <version> after building the bioagent bundle (cd bioagent && bun build src/main.ts --target=node --outdir=dist/server).

Layout

Component	Path
Binaries	/usr/lib/air-studio/bin/
Config	/etc/air-studio/
Agent backend	/opt/air-studio/bioagent/ (symlinked into /etc/air-studio/pai/bin)
Env scanner	/usr/local/sbin/air-studio-conda-r-versions + systemd timer
User data	~/.local/share/air-studio/
Listen port	7878 (configurable via www-port in /etc/air-studio/rserver.conf)

License

The RStudio Server code this project forks is AGPLv3 (see COPYING); modifications carry the same license. The toolbar logo is a placeholder — replace rstudio_2x_huaji.png if you redistribute.