Developer snapshots each agent round to a git ref and bisects to find regressions

On 2026-11-02 at 09:01, DEV Community's AI section published a walkthrough by developer xgabriel describing a workflow for tracking an Agent's changes with git. The approach starts with a Claude Code Stop hook that commits the Agent's work at the end of each turn into a private git ref, leaving the main branch, the staging area, and any stash untouched. Once every round is captured as its own snapshot, a script walks those snapshots through a bisect inside a temporary worktree, pinpointing exactly which turn introduced a regression and broke a test that had been passing before.

The scenario motivating the technique is a familiar one. The Agent reaches turn 20 and reports that the refund endpoint is finished, but running the suite shows test/refund.test.ts failing on a case that passed an hour earlier. Something changed somewhere between turn 1 and turn 20, and nothing in the repo makes that obvious, because the Agent never commits and the whole session's work sits in a single working-tree diff. A quick git log returns only the commit the session began from and nothing after it, so there is no commit-level trail to search.

By adding a Stop hook and a private ref per turn, the workflow restores the missing history in a way that does not interfere with normal git usage. The bisect script then checks out the recorded snapshots one at a time in a throwaway worktree, reruns the test at each step, and identifies the offending turn. The write-up presents this as a practical pattern for developers working with Claude Code, where long Agent sessions can silently accumulate changes across many turns. No broader adoption, tooling changes, or follow-up discussion beyond the article itself are documented.