Stop teaching your coding agent the same lesson every session.

rematch improves the agent's harness from the sessions you were going to run anyway, and keeps only what later sessions confirm. No reruns in interactive use; at most three model calls per reflection.

Start in 5 minutes How it decides

Install

scripts/install.sh
rematch doctor --text    # ready: true

cargo build --release
ln -s /path/to/rematch/adapters/claude-code ~/.claude/skills/rematch

cargo build --release
pi install "$PWD/adapters/pi"

Pi 1.4.x, from a Rematch checkout.

Add the checkout path to OpenCode’s plugin array:

{ "plugin": ["/path/to/rematch/adapters/opencode"] }

Build Rematch first. Tested with OpenCode 1.19.4.

ln -s ~/.rematch/skills ~/.agents/skills/rematch
rematch skills sync --harness codex --cwd "$PWD"

Run the sync before starting Codex in a repo.

The live adapters trace while you work; Codex uses post-hoc import. Reflection uses omp with model access. See installation and requirements before starting.

The loop

StartInstall, make your first correction, watch it get counted.

How it decidesThe loop, the evidence boundary, statuses and counters.

Day to dayCorrect, approve, create, improve, share, sync, report.

ReferenceEvery command, the trace schema, configuration, files.

What it will never do without you

Run a script or install a hook. Scripts are staged non-executable; hooks are compiled only after you approve them.
Change anything beyond your machine. Team and fork changes are branches and PRs; nothing is merged for you.
Write a repo’s AGENTS.md on your machine.
Call a change verified. A change that survives four matched sessions is kept, and its counts are shown raw.
Keep running after rematch off or REMATCH=0.