| name | hansel-and-gretel |
| description | Leave concise recoverable breadcrumbs through a long or multi-agent build so work can resume from evidence instead of memory. Use only when explicitly invoked to establish or inspect an operational checkpoint trail. |
Hansel and Gretel
Leave enough breadcrumbs to get home.
Breadcrumb
At each meaningful checkpoint record only operational facts needed to recover the work:
- current authoritative directive, packet, or objective;
- current step and completion state;
- concise user-visible decision rationale when a decision was made;
- files, artifacts, or interfaces changed;
- tests or evidence produced;
- unresolved questions, blockers, or known risks;
- next safe step;
- commit, hash, timestamp, or version when available.
Never record credentials, secrets, unnecessary personal data, or hidden chain-of-thought. If a fact is unavailable, mark it unknown.
Prefer append-only breadcrumbs so later work cannot silently rewrite earlier history.
Recovery
- Start from the latest trusted breadcrumb.
- Verify that its named artifacts and version still exist.
- Compare the current state with the recorded next step.
- If a gap or contradiction appears, stop and reconstruct that interval before continuing.
Output
BREADCRUMB
Authority: [directive or packet]
Step: [current step]
State: [complete/partial/blocked]
Changed: [artifacts]
Evidence: [tests/receipts]
Open: [questions/risks]
Next: [next safe step]
Version: [commit/hash/time or unknown]
Completion
Finish when another worker can resume from the breadcrumb without depending on conversational memory.