Entente checks agent work before it counts as done

Entente sits on top of Claude Code and Codex, requiring an agent that receives a task to accept a written contract or ask real questions before starting. Multi-agent handoffs usually end when the sending agent says it's done, and that self-report is what breaks down when the work is wrong. Entente removes that trust point by verifying the result itself, without relying on the agent's own account of it.

The enforcement comes after that handshake. A component called relayd reruns the contract's declared checks itself, inside the task's own worktree, and compares the real result against what the agent claimed. When the two disagree, relayd writes a delta repair naming only the criteria that failed. It attaches an explicit repair budget, so the agent fixes the gap instead of restarting from zero.

Every contract, check, human decision and repair gets appended to a JSONL log. A chain of agent handoffs leaves a written record, not a single pass-or-fail flag. The project is MIT licensed, needs Node 22+, and the repository went up on GitHub on September 4.

I haven't run it against a real multi-agent chain, so I can't say how the repair budget behaves when an agent can't clear the bar. The repo doesn't say what happens when a repair attempt exhausts its budget without passing.