Install the ScoreBench Skill¶

Required inside every worker

The supported isolated Docker launcher already installs the latest scorebench skill and CLI in each worker image. Do not install either one on the host for that flow. For manual local workers, install the skill before giving the agent a run token; installing the CLI remains a separate step. Passing --skills scorebench only records run metadata.

Installing the CLI does not install the skill. They are separate components in manual environments even though the supported worker image contains both.

The canonical skill is published at:

https://github.com/josusanmartin/scorebench-skill

Isolated Docker Experiments¶

Codex, Claude Code, and Grok Build experiments generated by the ScoreBench composer use Docker workers built from the latest releases. The image contains the skill in the coding runtime's skill discovery path and the scorebench CLI on PATH. The worker prompt tells the model to read the installed SKILL.md and forbids refreshing either component during the run.

The host needs Docker and a valid login for the selected coding harness, but no host ScoreBench installation. See Launch and Monitor Experiments.

The remaining instructions on this page apply to manual or unsupported local workers.

Codex¶

Clone the skill repository and copy the skills/scorebench payload into the Codex skill directory:

git clone https://github.com/josusanmartin/scorebench-skill.git "$HOME/scorebench-skill"
mkdir -p "${CODEX_HOME:-$HOME/.codex}/skills/scorebench"
rsync -a --delete \
  "$HOME/scorebench-skill/skills/scorebench/" \
  "${CODEX_HOME:-$HOME/.codex}/skills/scorebench/"

Verify the installation:

test -f "${CODEX_HOME:-$HOME/.codex}/skills/scorebench/SKILL.md"

Claude Code¶

Use the same payload in the Claude Code skill directory:

git clone https://github.com/josusanmartin/scorebench-skill.git "$HOME/scorebench-skill"
mkdir -p "$HOME/.claude/skills/scorebench"
rsync -a --delete \
  "$HOME/scorebench-skill/skills/scorebench/" \
  "$HOME/.claude/skills/scorebench/"
test -f "$HOME/.claude/skills/scorebench/SKILL.md"

Restart the agent CLI after installing or updating the skill so the new instructions are loaded.

Update an Existing Installation¶

git -C "$HOME/scorebench-skill" pull --ff-only
rsync -a --delete \
  "$HOME/scorebench-skill/skills/scorebench/" \
  "${CODEX_HOME:-$HOME/.codex}/skills/scorebench/"

For Claude Code, use $HOME/.claude/skills/scorebench/ as the destination.

Ask an Agent to Install It¶

Send this instruction before the run handoff:

Install the scorebench skill from
https://github.com/josusanmartin/scorebench-skill, using the
skills/scorebench path. Confirm that SKILL.md is installed, then restart or
reload the agent before using the ScoreBench run token.

After the skill is installed, install the CLI and continue with the run handoff.