Patchguard

The problem

Tools like Claude and Codex made generating code cheap, so the bottleneck moved. The new hard parts are trusting what changed, isolating a risky edit, and running more than one session without living in fear of the merge.

The approach

Patchguard is a layer around normal AI-assisted coding. It does not replace the model. You keep using Claude or Codex as you do now, and Patchguard makes the work isolated, reviewable, and easy to roll back. The first version is deliberately small: one active session in the repo at a time, isolated on its own Git branch, with a clear accept-or-reject review step.

Technical challenges

• Isolation and reversibility first. Before any orchestration, the goal was to make a single risky change trivial to undo. A lot of AI-workflow tooling reaches for parallelism too early; without clean rollback, more parallelism just means more ways to get hurt.
• A review surface that fits Git. Sessions map to branches with explicit review and finish steps, so the safety model is built on primitives I already trust.

Outcome

Phase 2 today: task coordination across sessions, branch-based isolation, and session review. The roadmap grows outward to worktree-based parallel sessions and, eventually, declarative orchestration through a Patchguardfile.

What I’d do differently

Nothing structural yet. The “smallest safe unit first” ordering has held up well as the foundation everything else rests on.