Patrol-native skill improvement

Build a Patrol-native self-learning skills and self-improvement harness

Reference paper: WikiSkill: Compiling Agent Experience into Persistent Knowledge for Skill Evolution

• arXiv: https://arxiv.org/abs/2608.27454 (link not exported)
• PDF: https://arxiv.org/pdf/2608.27454 (link not exported)

Context: Hive already owns two mature discovery controllers that should become the foundation of skill evolution rather than building a separate collector. Ordinary Patrol provides bounded snapshot-pinned evidence discovery, atomic evidence admission, semantic deduplication, deterministic routing, quotas, durable findings, verified repair loops, and review handoff. Architecture Patrol provides durable cross-boundary thesis jobs, rotating scopes, post-merge analysis, deduplication, explicit dispositions, routing, and audited history. Reuse those mechanisms to compile repeated agent experience into persistent llm-wiki knowledge and evaluated repo-local skill candidates.

Desired architecture:

• Ordinary Patrol is the local empirical signal lane: identify recurring execution failures, review findings, tool misuse, retry patterns, and successful corrective patterns from bounded eligible Hive artifacts.
• Architecture Patrol is the synthesis lane: cluster signals across tasks/components, distinguish code defects from instruction or workflow defects, and produce evidence-backed skill-evolution theses.
• llm-wiki is the provenance-backed knowledge substrate for accepted patterns, contradictions, freshness, scope, and rejected-candidate impact.
• A Hive-owned evolution controller compiles one thesis into one immutable repo-local skill candidate, runs blind held-out evaluation against active and fixed baselines, records results, and proposes an active-version transition.
• Activation is never automatic and never routes through agent-plugins. It requires explicit human approval through a Hive-owned compare-and-swap transition with smoke verification and rollback.
• Rejected candidates remain durable evidence and feed future patrol/architecture-patrol synthesis without contaminating held-out evaluation.

Acceptance criteria:

• Reuse shared Patrol and Architecture Patrol primitives for snapshots, source identity, evidence admission, deduplication, scoring or disposition, quotas, leases, durable jobs, routing, and audit records; do not fork equivalent infrastructure.
• Define typed signal, pattern/thesis, candidate, evaluation, decision, activation, and rollback contracts with stable IDs and digests.
• Add a target taxonomy that can route findings to code-fix, architecture thesis, knowledge-only update, or skill-evolution candidate without conflating these lanes.
• Ingest only bounded allowlisted Hive artifacts; exclude secrets, private conversations, raw unrelated transcripts, untracked or ignored files, and evaluator holdouts.
• Preserve train/development versus held-out isolation; proposer and worker cannot access hidden cases, rubrics, answers, or prior per-case results.
• Evaluate candidate versus current active version and a fixed baseline on immutable manifests; correctness and safety failures veto activation.
• Require explicit human approval for Hive-owned activation; no agent-plugins dependency or path.
• Retain accepted and rejected candidate identity, evidence, metrics, evaluator provenance, decision, impact, reason, and rollback lineage.
• Make cycles resumable and idempotent under crashes, retries, duplicate signals, default-branch movement, and concurrent daemon activity.
• Share the existing daily patrol ceiling and provider-native budget guards; expose clear status and hold reasons.
• Add end-to-end tests covering repeated-signal synthesis, deduplication, knowledge-only disposition, candidate proposal, blind evaluation, rejection, approval, activation, smoke-failure rollback, and recovery.
• Document how normal Patrol, Architecture Patrol, llm-wiki, and the skill-evolution harness interact, including ownership boundaries and operator controls.