KuraPre-release

Skills & Self-Improvement

Both follow one product rule: the agent proposes, the operator governs. No self-directed behavior outside the proposal/approval loop.

Agent-managed skills

A skill proposal is a governed memory asset (kind: skill):

  1. Propose — POST /v1/skills/proposals with name, description, body, and motivating evidence links (required — the validator refuses evidence-free proposals). Agent authorship is forced into Pending by the write policy, landing in the standard memory review queue.
  2. Review — the operator approves or rejects through the existing memory review (/v1/memory/assets/{id}/approve).
  3. Publish — POST /v1/skills/proposals/{id}/publish (approved proposals only; anything else is refused). Publication writes the SKILL.md bundle into <data_dir>/skills/, registers a Community catalog item whose version source permanently records memory:<assetId> (the provenance chain), and reloads the registry.

The runtime guard is structural: the skills registry only scans the skills directory, so pending or rejected proposals are never loadable.

Audited self-improvement

The closed loop the operator can audit and veto. Current target class: plugin-profile configuration values (e.g. session budgets, context budgets).

POST /v1/improvement/proposals
{
  "targetPlugin": "session-strategy",
  "configKey": "personalBudgetChars",
  "currentValue": 48000,
  "proposedValue": 64000,
  "predictedEffect": "fewer elisions in long sessions",
  "evidenceLinks": [{ "kind": "event", "id": "evt_ctx_123" }],
  "proposedBy": "agent"
}

Hard rules, enforced in code:

  • Evidence required — proposals without motivating evidence are refused.
  • Rate-bounded — max proposals per target per 24h window; the bound is operator configuration (self-improve plugin config), deliberately not agent-adjustable.
  • No change without a rollback path — apply snapshots the full prior profile into the proposal before atomically rewriting plugins.json; POST …/{id}/rollback restores it exactly.
  • Full audit chain — improvement.proposed → applied → kept/rolled_back are events; proposals persist as white-box JSON under <data_dir>/improvement/ and survive restarts.

Follow-up-evaluation with regression-triggered automatic rollback is the next slice on the roadmap.

Last updated: