# Reproduce the fresh-session recovery exercise

All names and project facts are synthetic.

1. Put README.md, decisions.md, proposal.md, status-old.md and prompt.txt in a dedicated folder. The README identifies the authoritative record.
2. Replace <trial-folder> in prompt.txt with that folder. Start a fresh session with no resume or continuation argument.
3. Run: hermes chat --safe-mode -t file --max-turns 8 --run-budget 120 --source tool --format stream-json --query-file prompt.txt --in <absolute-trial-folder>
4. Check result.json: acceptedDecisionId D02, supersededDecisionId D01, format JSON Lines, owner Jamie, next action a three-record import, evidenceFile decisions.md, healthVerifiedNow false. SQL stays unaccepted.
5. Hash source markdown before and after to ensure it was not changed.

The actual model/provider is recorded in receipt.json. Use your existing configured model; results can differ. Safe mode removes persistent-memory and automatic-rule assistance, so this proves explicit file recovery only. initial-result.json retains the first semantically correct output that failed the narrower downstream schema (filename anchor and null health field); the refined prompt was tested in a new session.