Skip to main content

What each kit command costs

reference · tech-lead · updated 2026-06-28 · source

A plain-language map of which budget every kit command spends, so you can tell at a glance what's safe to run a lot vs. what to run sparingly. Written after the GitHub rate-limit hit on 2026-06-24.

The four kinds of cost​

  • Free — runs only on your Mac (git + bash). No network, no bill.
  • GitHub — talks to GitHub, which caps how many calls you get per hour. Two flavors:
    • light — normal issue/PR calls; the hourly cap is generous.
    • heavy — anything that reads the whole project board. This is the cap that ran out on 2026-06-24.
  • Claude ($) — uses Claude (Anthropic). The only thing here that costs real money.
  • qwen (local model) — uses the small qwen model (local on your Mac, or your hosted endpoint) — not Claude. A separate, near-zero budget.

When you do X, it costs Y​

When you…Cost
Start / PR / merge / close a task (/kit-task-*)GitHub (light)
Read the board (/kit-task-sync)GitHub (heavy) — this is what drains the cap
Create / close an effort (/kit-effort-new, /kit-effort-close)GitHub (heavy — touches the board)
Open a worktree (/kit-task-start, /kit-effort-start)Free
Clean up (/kit-gc)GitHub (light)
Run several agents at once (orchestrate.sh, sub-agents, Workflow)Claude ($)
Summaries and the effort judge (/kit-digest, the telemetry scorer)qwen (local model)

The three things that matter​

  1. What drains GitHub = reading the board, repeatedly. /kit-task-sync and the board-touching effort commands are the heavy ones — don't run them in a tight loop.
  2. What costs money (Claude) = running several agents. The everyday task/effort commands are bash + GitHub — they cost nothing in Claude.
  3. The judge and the summaries use qwen, not Claude — a separate, near-zero model budget, so they never touch the Anthropic bill.

Why the board is the expensive part​

GitHub gives two hourly budgets: a large one for normal calls (issues, PRs) and a points-based one for "graph" queries. The kit reads the project board (300+ items) through the points-based budget — that's the only heavy call, and it lives in one helper (scripts/lib/gh-project.sh). Every board-touching command pulls it in, so a few of them back-to-back is what empties the cap.

The other always-on budget is the session-start skill catalog — see skill-catalog-audit.md for installed-vs-used and the prune list (#1031).