v0.1.3 — decision log
First public release line of breakerbox — hard spend caps and a kill-switch for the non-LLM actions your AI agent takes.
npm i -g @shopdevx/breakerbox && breakerbox init
Highlights
- Guardrail in Claude Code’s PreToolUse hook — reads every Bash command before it runs, prices it, and denies/asks if it breaches a cap. Zero deps, no build, ~48ms.
- Decision log (
.breakerbox/decisions.jsonl, new in 0.1.3) — one line per evaluated command: the command, the verdict, and whichfailModewas live. A crash is recorded asoutcome:"evaluation-error"; an absent line means the hook was never wired. - Fail-closed coverage (0.1.2) — a spawn test asserting
failMode:"closed"turns a crash into an explicitask, never a silent allow. - Belt and suspenders (0.1.2) — pair with
--allowedToolsso raw Bash is stripped when cloud access isn’t needed.
Honest limits
Spend guardrail, not a security sandbox (base64/eval evade it); only Bash is inspected today; prices are approximate list prices.
70 tests · MIT · https://shopdevx.github.io/breakerbox/
🤖 Generated with Claude Code