@deepseek-ai/dsh-evolution-curator

Deterministic skill lifecycle and recovery

Model Experience

Indirect model surface

What the model sees

@deepseek-ai/dsh-evolution-curator registers no direct prompt or tool schema itself. Model-visible effects are owned by the packages that consume this service.

Token effect

Zero direct token effect from this package; consumers add any model-visible tokens.

KV Cache effect

Independent of request-prefix construction. This package does not alter the assembled prompt or tool list.

Known Limitations and Deferred Work

• LLM nomination pass is advisory and disabled by default; deterministic lifecycle remains authoritative.
• Consolidation is control-plane only (/evolution consolidate <target> <source...>): no LLM pass proposes merge groups yet, and merged source bodies are appended verbatim rather than rewritten into a synthesized skill.

Recovery and consolidation

• archive never deletes: skills move to .archive/ with a .archive-reason marker.
• restore(name) (service) / /evolution skill restore <name> brings one archived skill back to the active root and resets its usage state.
• consolidate(target, sources) (service) / /evolution consolidate merges source bodies into the target, archives the sources with an absorbed-into marker, and folds their usage records into archived state. Both operations snapshot the full state first (pre-consolidate / pre-restore).
• restoreSnapshot (service) / /evolution restore rolls the FULL state back to the latest snapshot: active tree, usage/suppression sidecars, .archive/ and the curator state carried in the snapshot (curator-state.json), so the interval gate does not immediately re-fire after a rollback. The restore itself is undoable — the pre-rollback safety snapshot preserves the current tree plus its state.

Automatic scheduling

• autoStart (default true) arms an hourly interval check plus a deferred catch-up check bootGraceSeconds (default 10) after host boot. Both decide due-ness from the persisted lastRunAt, so a restart with an overdue schedule runs the first pass within the boot grace instead of waiting a full interval. bootGraceSeconds: 0 disables the deferral (not recommended: the check may run against a half-mounted host). All scheduling gates — interval, idle, first-run deferral, and the reentrancy guard — remain inside run().
• autoStart: false disables both automatic checks; /evolution curator run (manual, gate-skipping) still works.