another-harness Resume-Recover Environment Prototype

another-harness now contains its first executable resume_recover benchmark family.

Main architectural facts

• The new family is still synthetic, not yet a live recovery path.
• It is built explicitly around the Lean restart architecture:
  • WorkId.restart
  • ObservationScope
  • stream continuity vs attempt-local grounded completion
• The current two benchmark cases intentionally separate two truths:
  • a same-stream fresh-attempt recovery can succeed while remaining in_progress
  • an interrupted implementation can also be honestly finished to reviewed
• The recovering agent may update scoped implementation files, the work item, and the handoff, but it may not approve completion or mutate the evaluator artifact.
• Review tightened the family so missing handoffs score a failure instead of crashing, and empty required-substring contracts are rejected as vacuous.

Practical conclusion

The repo’s recovery story is no longer merely promised by the Atropos schema. It now exists as executable harness machinery, even if the live prepare-live and checkpoint/MCP integration cuts still remain to be earned.