watch.py admits heavy work only while PSI, swap-out rate and RAM forecast stay under the configured limits; config.default.json is the single default the install writes once (mode 600, no clobber) and every hook, admit and queue command falls back to. Browser runners and package-manager script forms count as heavy, preflight derives its probes from the light and heavy tool lists, and the shipped default carries no prune command, buildx path or host-specific key. Tests run once on unify/skills-20261114 after the merge (operator order 2026-11-14). 🎯 Fix: 85 % 💥 Break: 8 % 💸 Tokens: 99 % ⏱️ Velocity: – 🛡️ Security: 100 % 🏆 Score: 79 🚦 Gate: ⛔ fix 85 % < 90, break 8 % > 5 Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
3.6 KiB
name, description
| name | description |
|---|---|
| kvb-check-efficiency | Token cost of skill edits, CPU and RAM of code edits. |
💸 Tokens is the share of the base's resources a change uses: under 100 %
it saves, over 100 % it costs more. Paste the script's lines, never a
paragraph about them. <ref> is the commit the change
starts from: HEAD for uncommitted work, the merge base with the default
branch for a branch.
Skills: tokens
python3 ${CLAUDE_SKILL_DIR}/scripts/tokens.py --repo <skills repo> --base <ref> --transcripts ~/.claude/projects [--value <tokens/day> --evidence "<one line>"]
It compares every changed *.md under skills/, untracked ones included,
at <ref> with the working tree, and the always-loaded listing: all
description: words, counted by the repository's scripts/budget.py.
Tokens are an estimate, characters / 4. Changed files without any base text
print tokens new, which counts as over 100. Skills have no velocity.
A second line, 💰 value 1:<n> (<n>:1 when value < cost), weighs both
per day and never changes the exit code; pass its quotient or ? to
kvb-check-score --value.
- Cost: listing words delta x 1.33 x requests x 0.1 (cached) plus each
changed
SKILL.md's tokens delta x its invocations, per day of the last 7 days' transcripts (other runtimes: JSONL of that shape). A new skill takes its category's median, named in the line. - Value: tokens/day saved, derived by one
--evidenceline, e.g."3829 SendMessage wakes/day x 10 % avoided x 43k"; else?:?. - Measured: once shipped,
tokens.py value --repo <repo> --since <commit> --transcripts <dir>gives the median daily context tokens 7 days before minus after; pass it as--value.
Code: CPU and RAM
python3 ${CLAUDE_SKILL_DIR}/scripts/measure.py --base <ref> -- <command...>
Start it in the directory the command expects. It checks <ref> out as a
detached worktree under /tmp, runs the command there and in the working
tree from the same relative directory, compares CPU seconds and peak RSS,
and removes the worktree. Cost is the mean of the CPU and the RAM ratio.
It also measures wall time for kvb-check-speed.
- Runs: one per tree. Both values at most 100 end it; otherwise 4 more
per tree, alternating, and the line reports the mean of all 5
(
mean of 5). - Other load: a 1 s sample of system CPU busy before and after, and per
run the busy share beyond the command's own CPU time. Above 30 %, or
unreadable without
/proc/stat, the result is an estimate: time comes from the CPU-time ratio, both values carry~and the line ends in(estimated: other load <x> %). Pass the numbers without~tokvb-check-score; the exit code holds as for a measured result. - Command: the targeted tests of the change or a CLI call of the changed code, runnable on a clean checkout: the base worktree holds no untracked build output, virtualenv or dependency directory.
- Every run goes inside
${CLAUDE_SKILL_DIR}/../kvb-org-queue/scripts/queue.py hold --kind <kind> --task <register task> -- python3 <measure.py> ..., in the background, by thekvb-org-queueskill; a cancelled wait there also exits 3, but prints no 💸 line. - Report the run you made; never rerun to pick the best result.
Exit codes
0: at most 100.3: over 100, or new: pass the value tokvb-check-score; its gate asks the operator before the commit.2: usage error, or the command failed; the message names the tree, the exit code and the last stderr line. Fix the command, then measure.
--json prints the same figures as JSON.