Personal Project
Combustion
A framework for managing AI model delegation in Claude Code. Two levers, fuel and throttle, control cost-quality tradeoff and parallelism for multi-agent orchestration.
Combustion is me testing a simple idea: the expensive model should decide what to do, and cheaper models should do most of the doing. Opus is great at decomposition and review, but it costs more. Sonnet and Haiku are cheap and fast when the task is scoped well. The pricey part is figuring out the work, not typing it out.
I split control into two independent dials. Fuel sets cost and review strictness: lean ($, no review), tuned ($$, selective review), strict ($$$, full review), plus off/kill. Throttle sets speed and parallelism: coast (1-2), cruise (3-4), turbo (5-8). I keep them separate so I can run fast and cheap, or slow and strict, depending on risk.
To keep it usable in real work, I built a native Swift macOS menu bar app with hotkeys so I can switch fuel and throttle without breaking flow. Under the hood, Claude Code hooks (SessionStart, PreToolUse, SubagentStop), /fuel and /throttle skills, and a custom status line enforce policy and show model, context, fuel, throttle, and cost in real time.
Next up is broader tool support (Cursor, Windsurf) and stronger team governance defaults.