v0.23.0 — Personal Specialist Framework

What's New

• Personal specialist framework — user-owned, Agent Skills-compatible SKILL.md review methods live under the locally resolved wiki hub and are explicitly enabled per active topic.
• Deterministic management — initialize, create scaffolds, refresh indexes, validate packages, inspect provenance, and manage topic allowlists with llm-wiki specialist.
• Research integration — select up to three enabled specialists with repeatable --specialist flags, or force the baseline with --no-specialists; durable outputs record name, version, and SHA-256.
• Bounded safety model — v1 packages are instruction-only, cannot grant tools, and fail validation for scripts, binaries, symlinks, executable bits, or specialist-owned allowed-tools.
• No credential simulation — specialists encode evidence and review methods, not MBA, CFO, physician, PhD, legal, accounting, or fiduciary credentials; high-stakes work retains explicit human escalation.

Local/personal content boundary

This release ships the specialist infrastructure only. Personal HUB/.skills/ packages, topic allowlists, specialist definitions, and wiki-derived candidate reports are not bundled, tracked, or published. A release-archive audit confirmed none of those personal artifacts are present.

“Personal” means user-owned and stored in the selected hub. It does not mean encrypted or strictly local: a synced hub remains synced, and loaded instructions enter the selected model's context.

All deterministic plugin, structure, CLI, specialist, adapter, session, mirror-sync, token-budget, and clean-home Codex runtime tests pass. Claude plugin and marketplace manifests also pass strict validation.