fix: reproduce, then repair

fix: reproduce, then repair A workflow diagram generated by Archify. 01 / Host 02 / Model job: fix/from-ticket EX / Refusals and failures 04 / Write job: fix/propose + fix/report Ask DiagnoseThenFix Propose /fix comment · or an ai-generated issue · Host › Ask /fix comment or an ai-generated issue Pull request · closes the ticket · Host › Propose Pull request closes the ticket Write a reproducer · a test that fails on the bug · Model job: fix/from-ticket › DiagnoseThenFix Write a reproducer a test that fails on the bug Test, expect=fail · in a container · Model job: fix/from-ticket › DiagnoseThenFix · confirm red Test, expect=fail in a container confirm red Write the fix · shown the red test · Model job: fix/from-ticket › DiagnoseThenFix Write the fix shown the red test Test, expect=pass · reproducer plus fix · Model job: fix/from-ticket › Propose · confirm green Test, expect=pass reproducer plus fix confirm green fix.not_reproduced · the reproducer passed · Refusals and failures › DiagnoseThenFix fix.not_reproduced the reproducer passed fix.not_fixed · still red · Refusals and failures › DiagnoseThenFix fix.not_fixed still red fix/propose · title from the ticket · Write job: fix/propose + fix/report › Propose fix/propose title from the ticket passed red still red green: staged Legend Agent logic Policy External system

Red before green

  • • A reproducer that passes captured nothing, and the run says so
  • • Both suite runs happen in a container, on a throwaway worktree

Bounded

  • • Every turn is priced before it is made
  • • A failed attempt becomes an ai-generated issue, not a silent retry