01
Author
The next agent writes a small change on an isolated worktree branch.
Run local coding agents - Claude, Codex, Copilot, Gemini, Grok, agy - in a cross-audit loop on any git repo. One authors a small change; another reviews it; on agreement + green tests it merges. On disagreement, it revises. On stalemate, it asks you.
Agents never touch main. Only the engine merges, and only after the reviewer agrees and the test gate passes.
The next agent writes a small change on an isolated worktree branch.
A different agent reviews it read-only and returns AGREE or DISAGREE.
The repo's own test command runs. Red tests never merge - no exceptions.
Into orch/integration, with a persistent PR to main. GitHub owns main.
Diversity catches more. A Claude patch gets read by Codex - different model, different blind spots.
Agents run as your installed CLIs. Orchestration is local; agents connect to remote model APIs for inference.
Auto-detects npm, pytest, go, or make. No test command? It refuses to auto-merge and tells you.
Every merge-bound diff is pattern-scanned for secret reads, network calls, subprocess spawns, and guardrail edits - a check no reviewer prompt can talk its way past.
Killed mid-cycle? Committed work survives on its branch; the next run resumes from where it died.
Drive one repo with several cycles at once. File-overlap and post-merge re-tests guard the integration branch.
Before it prints "merged", orch confirms the commit is an ancestor of origin/main. No false successes.
Pin roles in orch.yml, or let the pool rotate the author each cycle.
orch initscaffold config + verify agent CLIs on PATHorch task "fix X"author + cross-audit + gate + merge - one cycleorch review <branch>audit an existing branch, no authoringorch pr <n>audit a GitHub PR, comment the verdict, optionally mergeorch dashboardlive cycle status, run history, success-rate metrics