A local harness for agent teams on Claude Code and Codex CLI: only what they prove counts as done. Waves and lanes split the work, reviewers from another model family judge it, councils keep the dissent that lost, typed decisions with Jev are measured before they are trusted, and approvals are bound to the bytes they approved. No model calls.
Executable compound practice loop: capture failures and decisions, surface them at the moment of action, and measure whether warnings prevent repeats. Bash CLI + MCP server.
