Repair in place
Repair is a mode of the implementer, not a new start. It uses the same worktree, the same scope and the same artifact. Use the findings together with the whole task acceptance contract.
Rules
- Continue the existing task. Do not repeat refinement or create a new plan for an implementation defect.
- Identify actual scope changes. Resolve missing facts from the repository; ask only when a plan defect changes the approved outcome or needs new authority.
- Fix in the lane's worktree, never by re-dispatch. A re-dispatch pays the scope-reading cost again with less budget.
- Check the gate's premise before you obey its remedy. The repair advice of a gate can be wrong.
Procedure
- Read the findings in the brief. When the cause is unclear or the failure recurs, run
root-causefirst. - Run the failing check in the worktree and read its output.
- Build any test fixture from the observed error, never from a suspected message. A fixture built from a guess tests a failure that never occurs.
- Write a regression test that fails.
- Fix the cause, then check every other call site of the same code.
- Run the check again and read the summary line.
- Commit and integrate in the existing worktree workflow. Loop-managed tasks may use their landing transition; direct observer repairs use ordinary Git and disclose any authorized hook bypass.
Do not pass a size gate by deleting comments or other text, and do not pass a test by loosening its assertion. Both make the gate green over the defect it exists to catch.