bench

Run the cc-proxy bench and show the user its output.

$1 selects the sub-command: grades (default), speed, or speed --report.

Execute:

# Resolve the plugin root. $CLAUDE_PLUGIN_ROOT is only injected in hook context,
# not into a slash-command's Bash, so fall back through the legacy PROXY_PATH
# pin, then the marketplace cache (newest version wins), then the dev repo.
# Same resolution as /cc-proxy:models — see that file for the two portability
# traps this form avoids (template substitution, and zsh word-splitting).
root=""
while IFS= read -r c; do
  [ -n "$c" ] || continue
  if [ -f "$c/scripts/bench-grades.js" ]; then root="$c"; break; fi
done <<EOF
${CLAUDE_PLUGIN_ROOT:-}
$([ -n "${PROXY_PATH:-}" ] && dirname "$(dirname "$PROXY_PATH")")
$(ls -d "$HOME"/.claude/plugins/cache/*/cc-proxy/*/ 2>/dev/null | sort -V -r | sed 's:/*$::')
$PWD
EOF
[ -n "$root" ] || { echo 'cc-proxy: cannot locate plugin root; run /cc-proxy:setup or /resume'; exit 1; }

# Trap 3 (this command's own): a slash-command body has NO real positional
# parameters. Claude Code substitutes the argument tokens TEXTUALLY before any
# shell runs, and it substitutes only $ARGUMENTS and $1 — $2 and $3 are left as
# the literal characters "$2"/"$3" for the shell to expand against an EMPTY
# argv, i.e. to nothing.
#
# Two bugs came from not knowing that, both verified against the real harness:
#   1. `shift; "$@"` forwarded nothing, silently dropping `--report` and turning
#      a read-only report into a full billed measurement run.
#   2. Reading `$1` then `$2`/`$3` was worse: for `speed --report` the harness
#      set $1 to `--report` (the LAST token, not the first), so `case "$1"`
#      matched no branch and fell through to the grades default — the wrong
#      sub-command entirely.
#
# So parse $ARGUMENTS — the whole argument string, the only token that carries
# every word — with `set --` word-splitting it into a real argv. The unquoted
# expansion is deliberate: these are shell-word arguments (`speed --report`),
# and the values are the user's own command line.
#
# `set -f` first, because unquoted word-splitting also GLOBS: measured in a
# directory holding two files, `bench speed *` split to three words
# (`speed aaa.txt bbb.txt`) instead of two. Splitting is what we want; pathname
# expansion is not, and the argv it builds is passed straight to a script.
set -f
set -- $ARGUMENTS
set +f
sub="${1:-grades}"
shift 2>/dev/null || true
case "$sub" in
  speed) node "$root/scripts/bench-speed.js" "$@" ;;
  grades) node "$root/scripts/bench-grades.js" ;;
  *)
    echo "cc-proxy: unknown sub-command '$sub' — expected 'grades', 'speed', or 'speed --report'" >&2
    exit 1 ;;
esac

Present the script's stdout verbatim — it is already formatted, and its footer states what each field means. Do not summarize, re-rank, or reword it.

Two things to keep straight if the user asks about the output:

• grade and score are different axes. grade is the model's position in its OWN vendor's line-up, read from the vendor's version numbering — that part needs no benchmark. score is benchlm's cross-vendor number, and a trailing ~ marks it as benchlm's estimate rather than a measurement (new models are usually estimated, which is exactly when the number deserves less trust). Never read one field off the other, and never let price lower a grade.
• speed measures the ROUTE, not the model. It never feeds grade. One ping is noise; speed --report gives median and p95 over the series, which is what makes route drift visible. A ⚠ spans >1 proxy build warning means the proxy binary changed mid-series and those numbers are not comparable.

grades needs network (benchlm.ai + OpenRouter) and writes ~/.claude/cc-proxy/grades.json. speed needs the proxy running and appends to ~/.claude/cc-proxy/speed.jsonl. On failure both say so and write nothing — a stale file is useful, a silently-empty one is a lie.