benchmark-implementlisted

# benchmark-implement Benchmark the harness process, not the subjective quality of the implemented product. Keep this skill as a thin wrapper around `$implement`: it prepares one isolated run, runs implementation in the coordinator's current session, gathers that run's records, requests one independent session evaluation, and builds `report.json`. ## Read Before Running Read: - [`references/contracts.md`](references/contracts.md) before creating or changing a benchmark case, generating a report, or comparing runs. - [`references/evaluator-rubric.md`](references/evaluator-rubric.md) before launching the evaluator or accepting `qualitative.json`. - `~/.codex/skills/implement/SKILL.md` completely before starting the implementation run, then follow every implement reference that applies. The Claude install substitutes the corresponding `~/.claude/skills/` path and `/implement` invocation. ## Artifacts Keep fixed inputs outside the harness bookkeeping namespace: ```text benchmarks/<case-id>/prd.md benchmarks/<case-id>/benchmark.json ``` Keep every run result under `agents/**`, which is excluded from judged implementation diffs: ```text agents/benchmarks/<case-id>/<run-id>/qualitative.json agents/benchmarks/<case-id>/<run-id>/report.json agents/benchmarks/<case-id>/<run-id>/run.json agents/benchmarks/<case-id>/<baseline>-vs-<candidate>.json ``` Do not copy the full transcript into the report. Store its path and hash, and cite stable session event identifiers in qualitati