benchmark

# benchmark — local regression / benchmark skill ## ⚡ Activation sentinel (READ FIRST) The FIRST RESPONSE you generate after reading this file in reaction to a "run benchmark" / "regression benchmark" / "benchmark clone" trigger MUST begin with the exact line below — no decoration, no preamble, no header above it: ``` [BENCHMARK-SKILL-ACTIVE v1 / skills/benchmark/SKILL.md] ``` Concretely: the response that comes right after the tool-result of your `Read(skills/benchmark/SKILL.md)`. Reading this file IS the start of execution; do not defer the sentinel to a later "more real" response, do not announce the rule in prose instead of emitting the token, do not prepend a step heading. The first line of that response is the token itself, and only the token. Anything else (Step 1 announcement, `bash skills/benchmark/scripts/setup.sh` call, status notes) goes below it on subsequent lines. Only that one response needs the sentinel — later responses in the same session do not need to repeat it. Empirically, LLMs drop a per-message prepend after a few turns even when the rule says "every response," so a single emission anchored to the post-read response is the reliable signal the maintainer checks. Edge case — tool-only first turn. If the response immediately after the SKILL.md Read would otherwise contain only a tool call with no text (e.g. you decide to dispatch `bash skills/benchmark/scripts/setup.sh` without saying anything), prefix that turn with the sentinel as a one-line text...