EqualifyEverything / benchmarks-ai-alt

The loop assumed Claude Code. Nothing about a round needs it: a round is one prompt and one turn against files and the web. - adapters/ holds one small file per harness, naming its command on a `# RUNS:` line. claude, codex and pi ship; template.sh is the fifth line of a new one. `--agent NAME` picks one, AGENT_CMD drives a CLI with no adapter, and with neither the first installed adapter is used and named. - Prompts go to the agent on standard input, with PROMPT_FILE, ROUND and ROLE in the environment, and `{prompt}` / `{prompt_file}` placeholders for CLIs that want an argument instead. - A harness with only a chat window can still run rounds: `--prompt seek` prints what to paste, `--apply N` promotes what the review accepted. Prompts are kept as round-NN-ROLE-prompt.md either way, so a hand-run round leaves the same audit trail as an automated one. - AGENTS.md is now the canonical rules file, read by more harnesses than CLAUDE.md, which points at it. The project README opens with a getting started walkthrough, because a list of script invocations is not enough to run this: check the harness with one line before spending a round on it, verify the machinery offline, read the directives that define what you are about to get, run a single round, then read the four files it wrote and check provenance, gold standards and reviewer bite by hand. Those checks are the part no tool here can do. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>