bough v0.2.7

What's Changed

• Open-source launch prep: green CI, commit hooks, AGENTS.md, docs by @andreylukin in #61

Full Changelog: v0.2.6...v0.2.7

bough v0.2.6

Full Changelog: v0.2.5...v0.2.6

bough v0.2.5

Full Changelog: v0.2.4...v0.2.5

bough v0.2.4

Full Changelog: v0.2.3...v0.2.4

bough v0.2.3

Full Changelog: v0.2.2...v0.2.3

bough v0.2.2

Full Changelog: v0.2.1...v0.2.2

bough v0.2.1

Full Changelog: v0.2.0...v0.2.1

bough v0.2.0

Full Changelog: v0.1.1...v0.2.0

bough 0.1.1

First public release.

bough is a coding agent whose only action is to write a program. Instead of emitting one tool call and waiting, the model writes JavaScript with real control flow and the harness runs it against your checkout — so a round is one program covering inspect, change and verify rather than five round-trips.

Install

brew tap andreylukin/bough https://github.com/andreylukin/bough
brew install bough

Or without Homebrew:

/bin/bash -c "$(curl -fsSL https://raw.githubusercontent.com/andreylukin/bough/main/install.sh)"

Both build from source and land on this tag. macOS or Linux; the first run takes a few minutes.

Before you run it

There is no isolation boundary. Programs run as you, with your full authority — filesystem, network, subprocesses. No sandbox, no egress proxy, no credential gating. That is deliberate rather than unfinished: the agent edits your real files because reviewing git diff and pushing with your own git is the delivery mechanism. Run it only where you would be comfortable running the code it writes.

What is in it

• One program per round, with view/patch as the editing idiom — the model names lines instead of quoting them, so code being edited never has to survive its own string escaping and a stale edit reports a conflict instead of clobbering.
• Command memory. Every shell command carries tags naming what it is for, and there is no untagged door. A later session finds what was already tried here.
• Note memory beside it, holding conclusions rather than commands: a note carries citations, never command strings, so a stale note gives wrong context but never a wrong incantation.
• History is a tree. Fork any turn; nothing is destructively rewritten.
• Four providers, routed by model id — Anthropic, OpenAI, OpenRouter, Cloudflare Workers AI. Every base URL is overridable, so an OpenAI-compatible server (Ollama, vLLM, LM Studio, a gateway) works through the OpenRouter path.
• Reads what you already have: AGENTS.md/CLAUDE.md per directory, ~/.claude/skills, and Claude Code's and Codex's configured hooks through adapters that ship on.

Documentation is in docs/; docs/spec.md and specs/ are authoritative for behaviour.

Apache-2.0.