Double Shot Latte

Why it's ranked here

The narrow scope is a strength. It targets a common break in multi-step coding flow, yet preserves pauses for questions, approvals, blockers, and completed work. Throttling, recursion protection, timeouts, and permissive failure behavior make the automation less likely to trap a session.

What's good

The judge receives structured output requirements and cannot use tools, which keeps its role constrained. Recent context is capped to control oversized transcripts. A failed command, timeout, missing transcript, or malformed response permits stopping. Cross-platform dispatch covers Unix and several Windows Bash locations.

Tradeoffs

Every stop attempt can invoke another Claude model, adding latency and model cost. Classification sees only the last four transcript lines, capped at 32 KB, so older intent may disappear. The default is conservative when uncertain. Windows still needs Bash, while every installation also requires jq and Claude Code.

How to use it well

Use it for autonomous, multi-file implementation where the agent frequently announces clear next steps. Keep the default judge for cheaper decisions, and change models only when classification quality warrants higher cost. It does not replace requirement gathering, human approval, error resolution, or project planning.

Observed

License

MIT License

Primary language

Bash shell scripting, with a cross-platform CMD wrapper

Install surface

Claude Code plugin installed from superpowers-marketplace

Interface

Claude Code Stop hook using JSON input and decision output

Platform support

Unix and Windows through Bash, including Git Bash, MSYS2, and Cygwin paths

Runtime prerequisites

Claude Code, jq, and Bash

Test structure

Repository includes shell test runners and JSON evaluation scenarios