Skip to content

Install

Install the binary, connect your harness, and check the connection.

One binary, ledger and library serve every harness; an adapter only forwards what its harness exposes. omp is the reference. What each harness can and cannot do: harness capabilities.

You need omp 18.12 and a Rust toolchain, in a checkout of this repo.

  1. Build and link:

    scripts/install.sh

    It builds the binary, links the extension into omp, and writes the shim ~/.local/bin/rematch.

  2. Check:

    rematch doctor --text

    The last line reads ready: true. Anything else is listed as a problem: line saying what to run.

The editor should be the strongest model you have, the classifier a fast one. Set them in ~/.omp/agent/config.yml:

modelRoles:
rematch-editor: <provider>/<strongest model> # falls back to slow → default
rematch-classifier: <provider>/<fast model> # falls back to smol → tiny → default

rematch doctor --text shows which models it resolved. Per-machine roles: model roles.

Import the sessions you already have, then reflect on the ones that carried a correction. Import is read-only and makes no model call.

rematch import --history
rematch reflect --history
  • import --history reads available omp, Claude Code, Codex, Pi and OpenCode history. Use --harness omp|claude-code|codex|pi|opencode to select one, or --skip-profile <name> to exclude an omp profile. Each adapter documents its history location and environment overrides.
  • reflect --history handles the newest 10 flagged sessions per run (--limit, --since 2026-10-01). Run it again for the next batch; a session is reflected once.
  • Each reflection follows the same rules as a live one: same rate limits, at most 3 model calls.