airb

An open-source, CLI-based programming agent for Rubyists.

🌌 Why use airb?

airb is an open‑source, CLI‑based programming agent for Rubyists. We built it to explore a clean, composable agent architecture grounded in cybernetics—specifically Stafford Beer's Viable System Model (VSM)—and to make a practical tool you can run in your terminal to read, list, and edit files with the help of modern LLMs.

In short:

• A new spine for agents: Operations, Coordination, Intelligence, Governance, Identity—recursive, inspectable, testable.
• A minimal, useful CLI that streams responses and uses structured tool calls (no fragile "JSON from text" parsing).
• A foundation to learn from and extend: add tools, swap models (OpenAI/Anthropic/Gemini), plug in UI/observability, grow sub‑agents.

If you like small objects, clear seams, and UNIXy ergonomics, airb is for you.

🌌 Who benefits from airb?

• Ruby developers who want a capable, hackable terminal agent that can actually work on a codebase.
• Framework/tool authors exploring agent design (capsules, tools, sub‑agents) with high cohesion & low coupling.
• Educators and teams who need a clear, auditable loop to reason about tool calling, streaming, and safety.
• Researchers playing with agent recursion (sub‑agents, tool‑as‑capsule), budget homeostasis, and observability.

🌌 What does airb do?

A CLI programming agent that:

• Streams assistant output to your terminal as it thinks.

• Uses native, structured tool calling across providers:

  • OpenAI "tools" (functions) — streaming + parallel tool calls
  • Anthropic tool_use / tool_result — streaming (input_json_delta)
  • Gemini function calling — non‑streaming MVP (streaming later)

• Ships with core programming tools (as capsules):

  • list_files(path?) — directory listing (dirs end with /)
  • read_file(path) — read UTF‑8 text files
  • edit_file(path, old_str, new_str) — replace/create with confirmation

• Runs on a VSM engine (via the vsm gem):

  • Operations — dispatches tool calls; concurrency per tool
  • Coordination — session floor control & turn lifecycle
  • Intelligence — LLM driver + conversation + streaming
  • Governance — workspace sandbox, confirmations, budgets
  • Identity — purpose/invariants & escalation hooks

• Provides observability from day one:

  • JSONL event ledger (.vsm.log.jsonl)
  • Optional web Lens (local SSE app) for live timelines

High‑level architecture

airb (top capsule)
├─ Identity        – name, invariants
├─ Governance      – sandbox, confirms, budgets
├─ Coordination    – floor control, turn end
├─ Intelligence    – driver(OpenAI/Anthropic/Gemini), streaming, tool loop
├─ Operations      – dispatch tools as child capsules (parallel)
│   ├─ list_files (tool capsule)
│   ├─ read_file  (tool capsule)
│   └─ edit_file  (tool capsule)
└─ Ports           – ChatTTY (CLI), Lens (web)

🌌 How to use it

Install

Requires Ruby 3.4+.

Add to your app or install globally:

# Using Bundler in a project
bundle add airb

# Or install gem globally
gem install airb

airb depends on the vsm gem (the agent runtime & drivers).

Configure a provider

Pick one provider and set env vars:

# OpenAI (streaming + tools)
export AIRB_PROVIDER=openai
export OPENAI_API_KEY=sk-...
export AIRB_MODEL=gpt-5-nano   # default if unset

# Anthropic (streaming + tool_use)
# export AIRB_PROVIDER=anthropic
# export ANTHROPIC_API_KEY=...
# export AIRB_MODEL=claude-sonnet-4-0   # default if unset

# Gemini (MVP: non-streaming tool calls)
# export AIRB_PROVIDER=gemini
# export GEMINI_API_KEY=...
# export AIRB_MODEL=gemini-2.5-flash    # default if unset

Quickstart

From the root of a Git repo:

airb

Sample session:

airb — chat (Ctrl-C to exit)
You: what's in this directory?
<streams…>
airb: README.md
      lib/
      spec/
      tmp/
You: open README.md
<streams…>
airb: (prints file contents)
You: replace the title with "Airb Demo"
<streams…>
confirm? Write to README.md? [y/N] y
<streams…>
airb: OK. Title updated.

Live visualizer (optional)

Start the local Lens web app (SSE):

VSM_LENS=1 airb
# Lens: http://127.0.0.1:9292

See live timeline & sessions: user messages, assistant deltas, tool calls/results, confirms, audits.

Configuration reference

Variable	Meaning	Default
AIRB_PROVIDER	openai \	anthropic \
AIRB_MODEL	Model name for chosen provider	see examples above
OPENAI_API_KEY	OpenAI auth	—
ANTHROPIC_API_KEY	Anthropic auth	—
GEMINI_API_KEY	Gemini auth	—
VSM_LENS	1 to enable web Lens	off
VSM_LENS_PORT	Lens port	9292
VSM_LENS_TOKEN	Optional access token (append ?token=...)	none

Workspace: airb auto‑detects repo root (git rev-parse). If not a repo, it uses Dir.pwd.

What happens on each turn?

1. You type text.
2. Intelligence appends it to the conversation and calls the provider driver.
3. The driver streams assistant text (assistant_delta).
4. If the model needs a tool, the driver emits tool_calls → Operations routes to the proper capsule.
5. The tool runs (in parallel, if multiple) and returns tool_result which is fed back to Intelligence.
6. The model produces a final assistant message; Coordination marks the turn complete.
7. Everything is emitted on the bus and logged; the Lens renders it live.

Provider behavior (at a glance)

Capability	OpenAI	Anthropic	Gemini (MVP)
Streaming text	✅ SSE	✅ SSE (text_delta)	➖ (planned)
Structured tool calls	✅ tools/tool_calls	✅ tool_use/tool_result	✅ functionCall/Response
Parallel tool calls	✅ supported	✅ supported	✅ supported
System prompt handling	in messages	header param (system)	in content / safety opts

airb normalizes these differences so your CLI experience is the same.

Advanced Usage

Add your own tool (as a capsule)

Create a class that inherits VSM::ToolCapsule, describe its schema, implement #run.

# lib/airb/tools/search_repo.rb
class SearchRepo < VSM::ToolCapsule
  tool_name "search_repo"
  tool_description "Search files for a regex under optional path"
  tool_schema({
    type: "object",
    properties: { path: {type:"string"}, pattern:{type:"string"} },
    required: ["pattern"]
  })

  # Optional: choose how it executes (fiber/thread/ractor/subprocess)
  def execution_mode = :thread

  def run(args)
    root = governance.send(:safe_path, args["path"] || ".")
    rx   = Regexp.new(args["pattern"])
    matches = Dir.glob("#{root}/**/*", File::FNM_DOTMATCH).
      select { |p| File.file?(p) }.
      filter_map do |file|
        lines = File.readlines(file, chomp:true, encoding:"UTF-8") rescue []
        hits  = lines.each_with_index.filter_map { |line,i| "#{file}:#{i+1}:#{line}" if rx.match?(line) }
        hits unless hits.empty?
      end
    matches.flatten.join("\n")
  end
end

Register it in your organism under Operations:

operations do
  capsule :search_repo, klass: SearchRepo
end

The Intelligence system automatically advertises it to the model as a structured tool (OpenAI/Anthropic/Gemini shapes).

Create a sub‑agent (recursive capsule)

When a "tool" needs multiple steps (plan → read → patch → verify), make it a full capsule with its own 5 systems (Operations/Coordination/Intelligence/Governance/Identity). Expose it as a tool by including VSM::ActsAsTool and implementing #run(args) that orchestrates internally, then returns a summary.

This keeps the parent simple while the sub‑agent stays cohesive and testable.

Concurrency & performance

• The runtime uses async fibers for orchestration and streaming.
• Each tool call runs in its own task; set execution_mode to :thread for CPU‑heavier work.
• Governance can add timeouts and semaphores to limit concurrent tool calls.

Safety & governance

• airb runs in a workspace sandbox (repo root or CWD).
• edit_file prompts for confirmation before writing.
• You can extend Governance to show diffs, enforce allowlists, or budget tokens/time.

Troubleshooting

• "No streaming" — Gemini driver is MVP (non‑streaming). Use OpenAI/Anthropic for streaming.
• "Path escapes workspace" — Governance blocked a write; run airb from the repo root or adjust logic.
• "No tool calls" — Ensure your provider key & model are set; some models require enabling tools.
• Lens shows nothing — Start with VSM_LENS=1, then open http://127.0.0.1:9292.

Table of Contents

• Why use airb?
• Who benefits from airb?
• What does airb do?
• How to use it
  • Install
  • Configure a provider
  • Quickstart
  • Live visualizer
  • Configuration reference
  • Provider behavior
• Advanced Usage
  • Add your own tool
  • Create a sub-agent
  • Concurrency & performance
  • Safety & governance
• Troubleshooting
• Roadmap
• Contributing
• License
• Acknowledgements

Roadmap

• Streaming for Gemini driver
• Diff previews & undo for edit_file
• MCP client/server ports (tool ecosystem)
• "Command mode" (airb -e "…") for one‑shot automation
• Rich Lens (search, replay, swimlanes, token counters)
• Additional built‑in capsules (planner, tester, editor)

Contributing

Bug reports, ideas, and PRs welcome!

• Please keep code SRP‑friendly, name things clearly, and favor composition over inheritance.
• Tests for drivers should include small fixture streams → expected events.

License

MIT (same as vsm), unless noted otherwise in subdirectories.

Acknowledgements

• Inspired by Stafford Beer's Viable System Model and the broader cybernetics community.
• Thanks to the Ruby OSS ecosystem for gems like async that make structured concurrency practical.
• Early discussions about good agent loops, tool use, and safety shaped this project.