v3.2.0 - Evidence-driven skill evolution

Optional experience-driven skill evolution with nine agent adapters, durable evidence, independent final testing and reviewed adoption.

Added

• Complete bounded evolution loop with persistent training evidence and patterns, multi-file proposals, retained failed attempts, validation-only selection, independent final testing and optional cross-agent transfer. Invalid model-authored patches are retained as rejections without aborting the remaining cycle.

• Nine host adapters (Claude, Codex, Cursor, Gemini, OpenCode, Pi, OMP, Hermes and OpenClaw) with neutral skill provisioning, durable interrupted tool traces and explicitly nullable usage; calibrated semantic judging, source-quotation checks, artifact tasks, and shared call/time/cost accounting across every role.

• Pre-request inference reservations in a bounded first-party API runner, with conservative model-context/output ceilings, Decimal accounting, no retry after uncertain responses, and expiring reviewed prices. CLI runners reject hard-USD configurations before inference.

• Ground semantic verdicts in exact factual-source or saved-artifact quotations, including changed artifacts without an explicit assertion. Artifact state does not establish the truth of newly authored claims.

• Task-cluster confidence intervals and paired randomization analysis, plus a 16-task public-project study covering real ingestion, index/log updates, graph compilation and hybrid-search execution.

• Portable skill/evidence bundles with SHA-256 verification; prepared runtime provisioning, disabled ONNX telemetry side effects and detached subprocess cleanup. Completed write-boundary violations retain evidence and score zero without ending the remaining isolated comparisons; runner/protocol failures still stop the run.

• Version 2 example suites and full-cycle regression tests. Fictional evaluation data lives under eval/evolution/pilot/, outside the installed skill, to avoid supplying answer keys to inference.

• Optional /wiki:learn workflow for immutable observed task experience, cited source/concept patterns, root-cause hypotheses, applicability and counterexamples.

• /wiki:evolve and stdlib wiki_evolve.py for evidence-linked whole-skill proposals and new skills, complete baseline/candidate snapshots, fixed-corpus paired evaluation, retained rejected attempts, explicit apply and rollback, whole-skill drift checks and interrupted-transition recovery.

• A provider-neutral JSON runner contract, optional authenticated Claude Code query adapter, and a 20-task fictional query pilot covering citations, current/historical decisions, contradictions, multi-source answers and abstention. No model improvement is claimed without a measured trial.

• Retained legacy version 1 regression gates; the new version 2 loop selects on validation only and reserves final tests for reporting. Both preserve failed attempts and reject malformed outputs and timeouts.

• Idempotent experience/pattern templates and experiment-archive guidance, an optional upgrade schema marker, end-to-end lifecycle tests and PR CI for Python and documentation.

Changed

• Credit the WikiSkill paper in the README and documentation, mapping the adopted methodology to implementation choices and published validation limits.
• Main plugin/skill version is 3.2.0: additive and backward-compatible, so no Markdown or CLI migration is required. Experience capture, skill editing and paid inference are never started by upgrade.
• Keep .evolution/ snapshots out of normal search, lint, statistics and graph extraction. Preserve searchable lessons and trial summaries in existing wiki page types.
• Paperclip companion paperclip-plugin-llm-wiki v0.5.3 adds matching archive exclusions in search, health and page reading. Its package README records publication status.
• Update command reference, agent workflows, site navigation, upgrade guide, contributor checks, manifest descriptions and machine-readable documentation. Handle documentation pages without a git modification timestamp during local builds.

Research basis: WikiSkill. The measured workflow study rejected both proposals; no quality gain was demonstrated. Five host integrations and the API budget path still need authenticated live validation.

Install or update the Claude marketplace plugin, or install skills/llm-wiki from this tag for other agents. Existing wiki upgrades are additive. Paperclip companion 0.5.3 is a separate npm publication; consult its registry entry for availability.