Upgrade to v3.2.0

v3.2.0 adds optional experience capture and evaluated skill improvements. It is a minor release because existing wikis, frontmatter and CLI flags remain compatible. It does not enable a background optimizer.

Update the installed plugin/skill, then run /wiki:upgrade against your configured wiki. The existing runtime setup remains mandatory. Upgrade adds .experience-template.json, .pattern-template.md and .evolution/README.md only if absent; custom files and SCHEMA.md are preserved. It proposes an optional ## Skill evolution schema section for review.

Use /wiki:learn for verified task experience, then /wiki:evolve for evidence-linked proposals and measured promotion. No evaluation or skill replacement occurs during upgrade. Back up .evolution/ as durable history, unlike the disposable .wiki-cache/. See Learning and skill evolution for the runner contract, pilot suite, inference budgets and rollback.

The Paperclip companion v0.5.3 excludes experiment snapshots from its browse/search/health surfaces. Update that companion separately when using it with an evolving wiki.

Upgrading from before v3.0

v3.0.0 replaces provider-backed embeddings with default local FastEmbed + sqlite-vec retrieval. Existing Markdown stays untouched; only the disposable vector index changes.

No content migration

No page or frontmatter format changed. The old embeddings.jsonl cache is ignored and can be deleted; /wiki:upgrade builds embeddings.sqlite locally and synchronizes every current section.

Why it's a major version

The default section search now initializes a local model and sqlite-vec index instead of staying lexical unless a provider is configured. The data boundary becomes strictly local and the old provider environment variables and approval flag are removed. Direct python wiki_search.py "<query>" --no-embed preserves dependency-free BM25.

The upgrade

1. Pull the current release. Update the plugin the same way you installed it: for Claude Code, re-run the marketplace install; for skill-only agents, re-run npx skills add.

2. Run the upgrade. In Claude Code:

   /wiki:upgrade

   Or directly, pass the installed script and the selected base directory separately. For a project wiki:

   python "${CLAUDE_PLUGIN_ROOT}/skills/llm-wiki/scripts/init_wiki.py" "<absolute-project-root>" --upgrade

   For the standard personal global layout:

   python "${CLAUDE_PLUGIN_ROOT}/skills/llm-wiki/scripts/init_wiki.py" "<absolute-home-directory>" --wiki-dir wiki --raw-dir wiki/raw --upgrade

   For a skill-only install, replace ${CLAUDE_PLUGIN_ROOT}/skills/llm-wiki with that agent's installed llm-wiki skill directory. Never use the plugin or skill directory as the base directory.

   Expected: Idempotent file operations leave existing pages untouched. Runtime setup then reports "status": "ready" with pinned dependency versions, model, page count, section count, and vector-index path before the command reports the missing SCHEMA marker.

3. Approve the SCHEMA.md merge. Add the local semantic backend lines from the current template to your existing ## Retrieval section. SCHEMA.md is co-evolved with you and is never modified silently.

Re-running the upgrade is idempotent: existing content stays unchanged, dependencies/model are verified, and only changed or missing vectors are embedded.

What changes for you

• Existing pages need no changes. Nothing canonical under wiki/ is rewritten.
• Section search is local hybrid by default. FastEmbed + sqlite-vec semantic ranking is fused with BM25. Run the script directly with Python and --no-embed for dependency-free lexical retrieval, or use --granularity page for whole-page ranking.
• Upgrade installs everything. The command resolves pinned FastEmbed, sqlite-vec, and PyYAML packages through uv, downloads the pinned model if absent, builds the parse cache, and embeds all current sections.
• wiki/.wiki-cache/ is regenerable. search-index.json and embeddings.sqlite are gitignored and safe to delete. Legacy embeddings.jsonl is ignored.
• No provider setup remains. No API key, endpoint, data-transfer approval, or per-query charge is involved.

WARNING

No wiki yet? There's nothing to upgrade. Run /wiki:init instead.