The next AI does not know what tests passed or failed
A handoff without verification state is dangerous. The next AI may assume the build passed, retry the wrong test, or ship a half-fix.
Why this hurts.
When test state is lost, the next session wastes time rediscovering whether it is implementing, debugging, or verifying.
How ShardStitch solves it.
Record the exact check and scope
Include verification state in the continuation packet.
Separate passed, failed, and unrun
Separate passed checks, failed checks, and checks not yet run.
Tie results to the current change
Tie failures back to changed files and next smallest action.
Rerun the smallest relevant check
Rerun the narrowest safe check against the current files; keep any broader unrun checks explicitly pending.
What the next AI receives.
- Commands run
- Pass/fail state
- Manual checks
- Next verification step
What stays out.
- Assumed success
- Unclear test claims
- Old failures already fixed
- Logs unrelated to the current task
Re-establish what has actually been verified
FAQ.
Does a handoff prove that a test passed?
No. A handoff can carry a recorded result, but the claim is only as reliable as its command output, scope, and freshness. Re-run the test when changes may have invalidated it.
What if I cannot find the old test output?
Mark the check unverified and rerun it if safe. If it cannot be rerun, state that limitation rather than claiming it passed.