Q Branch

Q Branch is four open-source gadgets for coding agents. This skill installs the ones you choose and explains how to drive them. Everything is MIT-licensed and public: https://github.com/grexican/qbranch

They share one idea: guarantee steps with hooks instead of longer prompts. A rule in CLAUDE.md or AGENTS.md costs tokens in every session and still gets skipped. A hook costs nothing until the moment it applies, and then it fires every time.

The gadgets

• headjack — Skills arrive when a task needs them. The library is indexed and searched per prompt instead of preloaded. This matters because large libraries lose their skill descriptions from the listing (measured on Claude Code 2.1.285: roughly 8 skills keep a description on Haiku, 105 on Sonnet), so a plain-language request can no longer be matched to a skill the agent technically has.
• middleout — One plan over your MCP servers: chained calls, one-to-many fan-out, automatic pagination, and only the fields you select come back.
• moneyball — Measures every subagent run from the transcripts (model, cost, and whether it actually worked), then graduates a kind of task to a cheaper model once the evidence supports it. It judges cost per completed task, so a cheap model that needs retries is not counted as cheap.
• groundhog — Mines transcripts for the lessons agents keep re-learning, keeps accepted ones in a committed ledger, and enforces them with a hook that blocks the known-bad command once, with the fix.

Honest limits

• groundhog needs session history to mine. A fresh machine has nothing to learn from yet.
• moneyball needs at least 10 runs of a task class before recommending anything, and its automatic model routing applies to Claude models; for other vendors it reports advice.
• headjack helps most with a large library. With a handful of skills, your agent's built-in preloading is fine, and the honest answer is to keep using it.
• middleout earns its place on per-item fan-out over chatty APIs. For a one-off call, a raw tool call is simpler. A plan runs under a single approval, so a fan-out of 100 items makes 100 calls; it is exactly as safe as the tools the plan names.

Requirements

Node 22.13+, or the native binary (Windows, macOS, Linux). Works with Claude Code, Cursor, Codex, Gemini CLI, Copilot and Windsurf. Where an agent cannot do something, that feature turns off rather than half-working, and each gadget's README lists exactly what works where.