At cycle start, a digest of the last N cycles — what shipped, what failed, what keeps recurring — is written into the run workspace (best-effort: a digest failure warns, never blocks). At the enforce stage the digest is injected into scout and triage prompts, positioned after the stable prompt prefix so LLM caching stays warm. Planning now sees history.
The canonical meta-defect of this repo: retrospectives kept writing precise diagnoses — "7th occurrence of this pattern" — into lesson files that no planning phase ever read. Signals were captured and then rotted. The chronicle campaign is the delivery mechanism:
| Cycle(s) | Slice | Story |
|---|---|---|
| 698 → 702 PASS | s2 digest writer (recurrence/digest.go) | The crash orphan: the Jul-11 power loss killed cycle 698 mid-scan; the recovered fleet re-picked the task and landed it in its first healthy wave |
| 725 PASS | s3 prompt wiring (cyclerun seeding, scout/triage injection) | A digest nobody reads changes nothing — this slice makes it planning input |
| queued | s6 escalation boundary · s7 historian · recurrence ledger | Repeated patterns should escalate weights automatically — the fleet-lane-provisioning-split FAIL ×4 is the standing test case |