01
Freeze after two failed repairs.
A third patch without a better theory is usually transcript drift wearing a hard hat. Stop and write the card.
The third blind repair attempt is usually where the agent starts lying to itself.
Use this when Claude Code, Codex, or another coding agent keeps patching without getting closer. The fix is a stop contract: write the failing repro, the invariant, and the exact proof command before touching one more file.
02
Keep the failure concrete.
"It still fails" is not enough. Capture the command, URL, log line, fixture, or visible behavior that proves the failure.
03
Name the invariant before the patch.
Most bad fixes change the output shape, auth boundary, data format, or user flow while claiming victory. Put the thing that must not move in writing.
04
Patch the smallest surface area.
If the agent wants to refactor a module to fix one failing case, it is probably avoiding the hard diagnosis. Smaller diff first.
05
Run the same proof path.
Do not let the agent invent a friendlier test after the fix. Re-run the command or workflow that exposed the bug unless there is a documented reason not to.
06
Write the stop contract.
Done, stuck, and ask-a-human need explicit definitions. Otherwise the agent will keep buying time with confident edits.