Token Sipping Mode

token-sipping-mode

**Delegate the legwork. Own the verdict.**

An operating mode for AI coding agents that must survive on a capped token budget: overnight runs, weekly rate windows, subscription tiers, long autonomous sessions. The expensive frontier model in your main loop spends tokens on exactly one thing, judgment. Everything that is cheap to specify gets fanned out to cheaper models, or to a second vendor's CLI that burns a separate budget entirely.

This repo is the whole kit:

Piece	What it does
SKILL.md	The operating mode itself: a routing table, a delegation contract, spend discipline rules, and the anti-pattern list
statusline/statusline.sh	A terminal statusline: model, context %, branch, thinking state, and rate-limit dot bars for your 5-hour and weekly windows
statusline/frugal/	The frugal meter: three small Python scripts that measure what delegation actually saved you, in dollars, live in the statusline

The statusline in action:

Fable 5 | 44% | my-project (feature/thing*) | thinking | frugal $603.41/$603.41 saved
current  ##OOOOOOOO 21%  10:20pm
weekly   ###OOOOOOO 26%  jul 30, 8:00am

That `frugal $603.41 saved` figure is not vibes. It is the difference between what your delegated subagent work would have cost at top-model pricing and what it actually cost at the tier it ran on, summed across the session and the lifetime of the metrics file.

The one law

Route each task DOWN to the cheapest tier that can do it correctly.

• Judgment stays home. Verdicts, synthesis, risk calls, anything that lands, posts, or gets graded: the main loop, on the best model you have.
• Verification goes mid-tier. Spec-clear multi-step work with checkable output: builds, test runs, diff analysis, structured research.
• Mechanical goes bottom-tier. Polls, greps, inventories, format checks.
• Second vendors are free capacity. A Codex CLI, a Gemini CLI, whatever you have: their budget is not your b