Repolane

Running several AI coding agents across many repositories at once leads to conflicting changes, mixed-up context, and agents straying outside their task into secrets, other repos or main branches.

Repolane previewRepolane preview
~10

parallel tasks across 5–6 isolated repositories

184

automated test cases guarding the rules engine

4

levels of scoped agent memory

The approach

An open-source (Apache 2.0) tool that gives each task its own git worktree and checks every agent action with a rules engine on Claude Code hooks, so agents can work in parallel safely. Designed for running many tasks in parallel: about 10 at once across 5–6 repositories, each in its own isolated worktree.

Architecture

  1. Workspaces: Git worktree per task, one clone per repo.
  2. Rules engine: Claude Code hooks check every agent action.
  3. Regression suite: 184 automated cases plus live self-checks.
  4. Memory & board: Scoped agent memory, local dashboard.

Capabilities

  • Designed a workspace model on git worktrees (one clone per repo, one isolated workspace per task) so several AI coding agents can work in parallel across repos without conflicts or mixed-up context.
  • Built a Python rules engine on Claude Code hooks that checks every agent action, including shell command chains and resolved file paths, to keep agents in scope, block access to secrets and protect main branches.
  • Guarded the rules against silent regressions with a 184-case automated test suite and live self-checks.
  • Designed scoped agent memory (preference, cross-repo, repo, task) so agents save what they learn at the right level, plus a zero-dependency local dashboard for tracking in-flight agent work across repos.
Open to opportunities

Building AI agents? Let's talk.

I'm looking for AI engineer roles where I can build agent systems end to end. Book a call or grab my resume, and I'll get back to you quickly.

Schedule a CallView Resume