Reproduce and fix issues
- Work in an isolated branch/worktree/sandbox when the host can provide one.
- Reproduce the exact reported behavior and save the smallest durable evidence.
- If reproduction fails, vary only evidence-backed environment assumptions; return the blocker instead of guessing.
- Trace root cause. If a cheap behavioral regression test exists, see it fail for the same reason.
- Make the smallest fix that explains the reproduction, then rerun the reproduction and targeted checks.
- Prepare a draft change/PR only when authorized. Post status back to the source thread/ticket with proof and remaining uncertainty.